(Ang Yang, Yin Shan) Intelligent Complex Adaptive

IGIP
I nt el l i gent Compl ex
Adapt i ve Syst ems
Ang Yang
CSIRO Land and Water, Australa
Yn Shan
Medcare Australa
Hershey New York
I GI PublI ShI nG
Acquisition Editor: Kristin Klinger

Senior Managing Editor: J ennifer Neidig
Managing Editor: Sara Reed
Development Editor: Kristin Roth
Copy Editor: Larissa Vinci
Typesetter: Amanda Appicello
Cover Design: Lisa Tosheff
Printed at: Yurchak Printing Inc.
Published in the United States of America by
IGI Publishing (an imprint of IGI Global)
701 E. Chocolate Avenue
Hershey PA 17033
Tel: 717-533-8845
Fax: 717-533-8661
E-mail: cust@igi-global.com
Web site: http://www.igi-global.com
and in the United Kingdom by
IGI Publishing (an imprint of IGI Global)
3 Henrietta Street
Covent Garden
London WC2E 8LU
Tel: 44 20 7240 0856
Fax: 44 20 7379 0609
Web site: http://www.eurospanonline.com
Copyright 2008 by IGI Global. All rights reserved. No part of this book may be reproduced in any form
or by any means, electronic or mechanical, including photocopying, without written permission from the
publisher.
Product or company names used in this book are for identifcation purposes only. Inclusion of the names of
the products or companies does not indicate a claim of ownership by IGI Global of the trademark or regis-
tered trademark.
Intelligent complex adaptive systems / Ang Yang and Yin Shan, editors.
p. cm.
Companion volume to: Applications of complex adaptive systems.
Summary: This book explores the foundation, history, and theory of intelligent adaptive systems, providing
a fundamental resource on topics such as the emergence of intelligent adaptive systems in social sciences,
biologically inspired artifcial social systems, sensory information processing, as well as the conceptual and
methodological issues and approaches to intelligent adaptive systems--Provided by publisher.
Includes bibliographical references and index.
ISBN-13: 978-1-59904-717-1 (hardcover)
ISBN-13: 978-1-59904-719-5 (ebook)
1. Functionalism (Social sciences) 2. System analysis. 3. Biocomplexity--Simulation methods. 4. Social
systems--Simulation methods. 5. Economics--Methodology. 6. Organizational sociology--Simulation meth-
ods. 7. Modularity (Engineering) 8. Modularity (Psychology) 9. Self-organizaing systems. 10. Adaptive
control systems. I. Yang, Ang. II. Shan, Yin.
HM484.I57 2008
300.11--dc22
2007032059
British Cataloguing in Publication Data
A Cataloguing in Publication record for this book is available from the British Library.
All work contributed to this book is new, previously-unpublished material. The views expressed in this book
are those of the authors, but not necessarily of the publisher.
I nt el l i gent Compl ex
Adapt i ve Syst ems
Tabl e of Cont ent s
Foreword.................................................................................................vi
Preface.................................................................................................. viii
Section.I
General.Theories
Chapter.I
From.Reductive.to.Robust:.Seeking.the.Core.of.Complex.
Adaptive.Systems.Theory.......................................................................1
Steven E. Wallis, Independent Consultant, USA
Chapter.II
Method.of.Systems.Potential.as.Top-Bottom.Technique.of.
the.Complex.Adaptive.Systems.Modelling.........................................26
Grigorii S. Pushnoi, St. Petersburg, Russia
Gordon L. Bonser, California, USA
v
Section.II
Important.Concepts
Chapter.III
Modularity.and.Complex.Adaptive.Systems......................................75
David Conforth, University of NSW, Australia
David G. Green, Monash University, Australia
Chapter.IV
Concept and Defnition of Complexity..............................................105
Russell K. Standish, UNSW, Australia
Section.III
Computing.Perspectives
Chapter.V
Emergence.of.Creativity:.A.Simulation.Approach...........................126
Hrafn Thorri Thrisson, Reykjavk University, Iceland
Chapter.VI
Solving.the.Sensory.Information.Bottleneck.to.Central.
Processing.in.Adaptive.Systems.........................................................159
Thomy Nilsson, University of Prince Edward Island, Canada
Chapter.VII
Complexity,.Information,.and.Robustness:.The.Role.of.
Information.Barriers.in.Boolean.Networks..................................187
Kurt A. Richardson, ISCE Research, USA
Chapter.VIII
Emergent.Specialization.in.Biologically.Inspired.Collective.
Behavior.Systems................................................................................215
G. S. Nitschke, Vrije Universiteit, The Netherlands
M. C. Schut, Vrije Universiteit, The Netherlands
A. E. Eiben, Vrije Universiteit, The Netherlands
v
Section.IV
Social.Science.Perspectives
Chapter.IX
Emergence.in.Agent-Based.Computational.Social.Science:.
Conceptual,.Formal,.and.Diagrammatic.Analysis...........................255
Jean Louis Dessalles, ENST, Paris, France
Jacques Ferber, LIRMM, CNRS, & University of Montpellier,
France
Denis Phan, GEMAS, CNRS, and University of Paris IV
Sorbonne, & CREM, CNRS, and University of Rennes I,
France
Chapter.X
Ontological Refections on Peace and War.......................................300
Hayward R. Alker, University of Southern California &
Watson Institute, Brown University, USA
Chapter.XI
The.Allocation.of.Complexity.in.Economic.Systems.......................331
Jason Potts, University of Queensland & Queensland
University of Technology, Australia
KateMorrison,VolterraPacifcPty.Ltd.,Australia
Joseph Clark, University of Queensland & Suncorp Pty. Ltd.,
Australia
About.the.Contributors......................................................................351
Index.....................................................................................................358
v
For ewor d
I dont believe in the existence of a complex systems theory as such and, so
far, Im still referring to complex systems science (CSS) in order to describe
my research endeavours. In my view, the latter is constituted, up until now,
by a bundle of loosely connected methods and theories aiming to observe
from contrasted standpointsthese fascinating objects of research called
complex adaptive systems. Nearly 40 years after Von Bertalanffys General
System Theory (1968) and J acques Monods Chance and Necessity (1971),
it is fair to look back and to try to assess how much remains to be said about
these complex adaptive systems. After all, Prigogines Order out of Chaos
(1984) already demonstrated that future wasnt entirely predictable in a his-
tory-contingent world. Nearly at the same period, Maturana and Varelas
Tree of Knowledge (1987) questioned the closure of biological systems and
proposed a challenging theory of autopoieitic systems, oddly left aside by
CSSs mainstream research. Later on, Hollands Hidden Order (1996) set out
the terminology associated with and the characteristics of complex adaptive
systems, still in use nowadays. More recently, Wattss Six Degrees (2004)
epitomized current assumptions of network theorists asserting that a systems
structure and organizationmost of the timedictate its functional proper-
ties. What remains from these infuential contributions are a heterogeneous
corpus of partly conficting theories and a disparate set of tools and methods.
Furthermore, too often complex systems science lends itself to criticism when
it trades its artifcial complex adaptive systems for natural (i.e., actual) ones.
v
Computer-based simulations, regardless of their expected accuracy, arent
the reality, there are just metaphoric representations; the world as it might
be, not the world as it is according to Holland himself.
So, yes, much remains to be said about complex adaptive systems (CAS).
Altogether, we need to better our understanding of natural CAS and to im-
prove analytical capacities of artifcial CAS. Both aspects need to be dealt
with cautiously in order to avoid ill-fated circularities that have sometimes
characterized research out of in-vitro simulations or artifcial society experi-
ments. It is indeed an understated challenge to design a computer metaphor
that describes a given reality independently of the hypothetical processes
to be tested. Flawed designs often result in logical tautologies whereby the
model always verifes the assumptions. Another challenge consists in the
reconciliation between system-wide and individual-centred representations
of CAS. This task is anything but trivial as technical limitations and epis-
temological differences have contributed to the divide. Technically, latest
hybrid simulation platforms provide the means to couple agent-based mod-
elling with dynamical systems modelling or network-oriented simulations.
But epistemological differences on internationality, for example, need to be
dealt with in a same way biology has progressively dealt with the tension
between Lamarckism and Darwinism on evolution.
In this context, the present ouvrage comes at its time. The carefully se-
lected chapters cover the latest theoretical developments on natural CAS
and innovative ways to improve the analytical capacities of artifcial CAS.
Traditional concepts of complex systems science are re-visited: what is emer-
gence? Can we explain the emergence of creativity in natural CAS? How
does emergent specialization improve artifcial CASs design? Likewise,
essential characteristics of social CAS are scrutinized: How do information
fows infuence the complexity of social systems? Can we propose a robust
ontological foundation for social simulations? Finally, this book invites us
into an interdisciplinary journey through biological evolution, neo-classical
economics, system thinking and social sciences, using CSS as its Arians
thread. Intelligent complex adaptive systems (ICAS) will emerge from this
interdisciplinary cross-fertilization combined with technological advances.
They will provide powerful analytical capacities, supported by a reunifed
and holistic vision on complex adaptive systems. They will help us to build
what I will have to call, fnally, a complex systems theory.
Pascal Perez
AssociateProfessor,ResearchSchoolofPacifc&AsianStudies,Australian
National University
v
Pr ef ac e
Our world is a large, integrated system of systems. These systems, whether
they are ecological, social, or fnancial, are complex and constantly adapt to
their environment. Many of them are essential for our very existence. Be-
ing so complex, and because of the intensive interactions among the system
components, they cannot be fully understood by isolating their components or
applying simple cause and effect reasoning. These systems, however, can be
examined by looking for patterns within their behaviour. Intelligent complex
adaptive systems (ICAS) research uses systemic inquiry to build multi-dis-
ciplinary representations of reality to study such complex systems.
Because the use of ICAS is prevalent across a number of disciplines, papers
describing ICAS theory and applications are scattered through different
journals and conference proceedings. It is, therefore, important to have a
book that broadly covers the state-of-art in this highly evolving area. There
has been a strong interest among researchers regarding the publication of
this book. Forty-nine submissions were received. All papers went through
rigid peer review by at least three reviewers and only 23 were accepted for
publication, an acceptance rate of just under 50%. Because of size constraints,
these papers are published two volumes. This book focuses on the theoreti-
cal side of ICAS while its sister book Applications of Intelligent Complex
Adaptive Systems emphasises the techniques and applications. These two
volumes cover a broad spectrum of ICAS research from discussion of general
theory and foundations to more practical studies of ICAS in various artifcial
x
and natural systems. It is important to highlight that a signifcant portion of
contributions come from the social sciences. This will, we believe, provide
readers of these books with extremely valuable diverse views of ICAS, and
also clearly demonstrates the wide applicability of ICAS theories.
Intelligent.Complex.Adaptive.Systems
The study of ICAS draws richly from foundations in several disciplines,
perhaps explaining in part why ICAS research is so active and productive.
These diverse felds that contributed to the formation of ICAS included the
genetic algorithm (Holland, 1975) and cellular automata (Gardner, 1970,
von Neumann, 1966) in computer sciences, evolution and predator-prey
models (Lotka, 1925) in biology, and game theory (von Neumann & Mor-
genstern, 1944) in economics.
Researchers of ICAS are interested in various questions, but these can be
summarised as to how to describe complex systems, and how to describe the
interactions within these systems that give rise to patterns. Thus, although
researchers from different backgrounds may have very different approaches
to the study of ICAS, it is the unique properties of ICAS systems, such as
nonlinearity, emergence, adaptivity and modularity that form the centre of
inquiries. Many of these properties will be thoroughly explored in these two
volumes. It is the complexity of ICAS systems which means that although a
variety of techniques which have been employed to study ICAS, computer
simulations have become important and widely used. These simulations
involve several important computing techniques that may interest readers
of these books.
Evolutionary computation (EC) is a highly active feld of research
inspired by natural evolution. Essentially, EC models the dynamics
of a population of distinctive entities such as chromosomes in genetic
algorithms or programs in genetic programming. Thus, while EC has
been used as a simplifed model to study ICAS, it is also an ICAS itself
having wide applicability for solving scientifc and engineering prob-
lems.
Cellular automata (CA), and related techniques such as Boolean net-
works, are common techniques in ICAS. The behaviour of entities that
respond to the environment is defned as rules or other forms. Each
x
entity can interact with adjacent ones. The topology of adjacency can be
defned in various ways depending on the focus of the research. CA and
related techniques have been widely used to study important properties
of ICAS such as emergence.
Multi-agent systems (MAS) are systems composed of several autono-
mous agents. These agents may use a belief-desire-intention model or
other mechanisms to guide their behaviour, respond to the environment,
or communicate and interact with other agents. The concept of MAS
model can be directly applied to study a number of ICAS systems. More
often, a computer simulation of MAS is used to understand correspond-
ing ICAS.
ICAS research has applications across numerous disciplines. As we are sur-
rounded by complex systems, and indeed are ourselves a complex system,
applications are everywhere. In this preface, we have no intention of provid-
ing a compete list of applications of ICAS, although some of the chapters do
survey ICAS applications in a particular feld, but we do wish to highlight
the following subjects that are covered by this book and its sister volume.
Because human society is a complex system, comprising a large number of
autonomous individuals and entities that are connected by various layers of
networks, it has been the one of the major felds of applications of ICAS
research. As explained in a number of excellent chapters, signifcant research
has been conducted into how disease, information, belief, language, and in-
novation propagate and diffuse in society.
Economics and fnance are also the focuses of applied ICAS research. The
economic and fnancial interactions among the entities of modern society,
either at individual or institutional level, are vital to its existence. ICAS has
been used to study these interactions and to understand the dynamics that
underpin them.
Management can also been understood and further explored with ICAS
concepts and methodologies that provide both a novel perspective and an
exciting new set of tools. Besides applications to general management, these
two books also have chapters dedicated to specifc management applications
such as military transformation.
And fnally, ICAS has been widely used in science and engineering. Complex
systems exist almost everywhere in the natural world, from the complex
dynamics of the weather to important ecological systems. ICAS plays an
important role in understanding these systems. Furthermore, it is well known
x
that the robustness and reliability of an ICAS system is partially due to the
fact that there is usually no centralised control system. This idea has been
explored in solving engineering problems.
Audience
Researchers working in the feld of ICAS and related felds such as machine
learning, artifcial intelligence, multi-agent systems, economy, fnance, man-
agement, international relations, and other social sciences should fnd this
book an indispensable state-of-art reference. Because of its comprehensive
coverage, the book can also be used as complementary readings at the post-
graduate level.
Organisation
The diversity of backgrounds within ICAS research provides the deep well
of intellectual resources, which has allowed ICAS research to continue thriv-
ing. This diversity of backgrounds leads to a blurring of boundaries when
categorising ICAS studies, a feature that should be seen as one of the unique
characteristics of this feld of research and one that need not be regarded as a
problem. In organising a book, however, divisions are necessary. Thus, our
organisation of this book into four sections on general theories, importance
concepts, and perspectives in computing and the social sciences is only meant
to provide the audience with a simple reference to make the book more ac-
cessible. The interdisciplinary nature of ICAS means that many articles might
ft into multiple sections.
Two chapters in the general theories section seek the core of complex adaptive
systems (CAS) and provide an alternative top-down method called method
of systems potential rather than conventional agent-based modelling to
study CAS. The frst chapter by Wallis seeks to identify the core of CAS
theory. To achieve this, it introduces innovative methods for measuring and
advancing the validity of a theory by understanding the structure of theory.
Two studies of CAS theory are presented. These show how the outer belt
of loosely connected concepts support the evolution of a theory while, in
contrast, the robust core of a theory, consisting of co-causal propositions,
x
supports validity and testability. The tools presented in this chapter may be
used to support the purposeful evolution of theory by improving the validity
of ICAS theory.
In the second chapter by Pushnoi et al. emergent properties of CAS are
explored by means of agent-based modelling (ABM) and compared to
results from method of systems potential (MSP) modelling. MSP describes
CAS as a holistic system whereas ABM-methodology considers CAS as set
of interacting agents. The authors argue that MSP is a top-down approach,
which supplements the bottom-up modelling by ABM. Both ABM and MSP
exhibit similar macroscopic properties such as punctuated equilibrium, sud-
den jumps in macro-indices, cyclical dynamics, superposition of determin-
istic and stochastic patterns in dynamics, fractal properties of structure and
dynamics, and SOC-phenomenon. ABM demonstrates these properties via
simulations of the different models whereas MSP derives these phenomena
analytically.
In the second section of this book on important concepts, two chapters seek
to understand modularity, hierarchy, complexity, and emergence in the con-
text of ICAS. Cornforth et al. provide an in-depth discussion of modularity,
ubiquitous in CAS. Modules are clusters of components that interact with
their environment as a single unit and, when they occur at different levels,
form a hierarchy. Modularity occurs widely in natural and artifcial systems,
in the latter where it is used to simplify design, provide fault tolerance and
solve diffcult problems by decomposition. Modular and hierarchic structures
simplify analysis of complex systems by reducing long-range connections,
thus constraining groups of components to act as a single component. Net-
work theory, widely used to understand modularity and hierarchy, can make
predictions about certain properties of systems such as the effects of critical
phase changes in connectivity.
In the study by Standish, investigation of important concepts in CAS is
continued. The term complexity is used informally both as a quality and as
a quantity. As a quality, complexity relates to our ability to understand a
system or objectwe understand simple systems, but not complex ones. On
another level, complexity is used as a quantity when we talk about something
being more complicated than another. This chapter explores the formalisa-
tion of both meanings of complexity during the latter half of the twentieth
century.
The third section of this book features four chapters on the use of modern
computing techniques to study the emergence of creativity, emergent spe-
cialisation, information bottleneck to central processing in adaptive systems
x
and the role of barriers to information fows in the robustness of complex
systems.
Creativity has been a diffcult concept to defne and its exact relationship
with intelligence remains to be explained. In the frst of the four chapters on
computing techniques, Thrisson presents a theory of natural creativity and
its relation to certain features of intelligence and cognitive faculties. To test
the theory, the author employs simulated worlds of varying complexity that
are inhabited by creatures with a genetically evolving mental model. Plan-
making strategies are compared between creatures in each world. This shows
that creative behaviours are governed by the worlds structural coherence and
complexity. The theoretical framework presented in this chapter may serve
as a foundation and tool to improve our understanding of natural creativity
and to help develop creative artifcially intelligent systems.
Nilsson contributes the second study on computing perspectives, considering
sensory information bottlenecks in adaptive systems. Such bottlenecks are
an inevitable consequence when a complex system adapts by increasing its
information input. Input and output bottlenecks are due to geometrical limits
that arise because the area available for connections from an external surface
always exceeds the area available for connections to an internal surface.
Processing of the additional information faces an internal bottleneck. As
more elements increase the size of a processor, its interface surface increases
more slowly than its volume. Such bottlenecks had to be overcome before
complex life forms could evolve. Based on mapping studies, it is generally
agreed that sensory inputs to the brain are organized as convergent-divergent
networks. However, no one has previously explained how such networks can
conserve the location and magnitude of any particular stimulus. The solution
to a convergent-divergent network that overcomes bottleneck problems turns
out to be surprisingly simple and yet restricted.
In the current information age, a high premium is placed on the widespread
availability of information, and access to as much information as possible is
often cited as a key to making effective decisions. While it would be fool-
ish to deny the central role that information fow has in effective decision
making processes, in the chapter by Richardson, the equally important role
of barriers to information fows in the robustness of complex systems are
explored. The analysis demonstrates that, for simple Boolean networks, a
complex systems ability to flter out certain information fows is essential
if the system is not to be beholden to every external signal. The reduction of
information is as important as the availability of information.
xv
Specialisation is observable in many complex adaptive systems and is thought
to be a fundamental mechanism for achieving optimal effciency. In the fnal
chapter on computing perspectives, Nitschke et al. present a survey of col-
lective behaviour systems designed using biologically inspired principles
in which specialization emerges as a result of system dynamics, and where
emergent specialization is used as a problem solver or means to increase
task performance. The authors argue for developing design methodologies
and principles that facilitate emergent specialization in collective behaviour
systems.
In this books fnal section, three chapters provide insight into emergence
in multi-agent systems in the social sciences, and the application of ICAS
theories in international relations and economic systems. Dessalles et al. pro-
vide a survey of concept of emergence from both a conceptual and a formal
perspective, and discuss the notions of downward/backward causation and
weak/strong emergence. They pay particular attention to the formal defni-
tions introduced by Mller and Bonabeau, which are operative in multi-agent
frameworks and are derived from both cognitive and social points of view.
In the second social sciences chapter, Alker provides a study on ontological
refections on peace and war. Responding to a question by Hiroharu Seki
about Hiroshima ontologies, the author reviews thinking about the ontological
primitives appropriate for event-data making, accessing high-performance
knowledge bases, and modelling ICAS used by researchers on war and peace.
It cautions against Cliocide, defned as of the silencing or symbolic
killing of collective historical-political or historical-disciplinary identities
and identifying practices by historical or discipline defcient scientifc
coding practices. He proposes that more intelligent multi-agent models in
the complex, adaptive systems tradition of the Santa Fe Institute should
include the socially shared memories of nations and international societies,
including their identity-redefning traumas and their relational/migrational/
ecological histories of community-building success and failure. Historicity
in an ontologically distinctive sense of the time ordered self-understandings
of a continuing human society is still a challenge for the computationally
oriented literature on war and peace.
In the fnal chapter of this book, Potts et al. discuss a classic allocation
problem. The substitution relation between two primary carriers of complex
rulesagents and institutionsis a function of the relative costs of embed-
ding rules in these carriers, all subject to the constraint of maintaining overall
system complexity. This generic model is called the allocation of complex-
ity, which they propose as a bridge between neoclassical and complexity
economics.
xv
Conclusion
This book and its sister volume bring together prominent ICAS researchers
from around the globe who provide us with a valuable diverse set of views
on ICAS. Their work covers a wide spectrum of cutting-edge ICAS research,
from theory to applications in various felds such as computing and social
sciences, and provides both comprehensive surveys on some topics and
in-depth discussions on others. This offers us a glimpse of the rapidly pro-
gressing and extremely active feld that is ICAS research. More importantly,
because of the interdisciplinary background of the contributors, these books
should facilitate communications between researchers from these different
felds and thus help to further enhance ICAS research. Thus, we hope that
these books may help to raise the profle of the contribution that complex
adaptive systems can make toward better understanding of the various criti-
cal systems around us. In doing so, this work should encourage both further
research into this area and also the practical implementation of the results
derived from this area.
References
Gardner, M. (1970). Mathematical games: The fantastic combinations of John
Conways new solitaire game life. ScientifcAmerican,223(October
1970), 120-123.
Holland, J . H. (1975). Adaptation in natural and artifcial systems. Ann
Arbor, MI: University of Michigan Press.
Lotka, A. J . (1925). The elements of physical biology. Baltimore, MD: Wil-
liams & Williams Co.
Von Neumann, J . (1966). Theory of self-reproducing automata. In A. W.
Burks (ed.), Lectures on the theory and organization of complicated
automata. Champaign, IL: University of Illinois Press.
Von Neumann, J . & Morgenstern, O. (1944). Theory of games and economic
behaviour. Princeton, NJ : Princeton University Press.
Ying Shan, Ang Yang
Canberra, Australia
June2007
xv
Section I
General Theories
From Reductve to Robust
Copyright 2008, IGI Global. Copying or distributing in print or electronic forms without written permission
of IGI Global is prohibited.
Chapter.I
From.Reductive.to.Robust:
Seeking.the.Core.of.Complex.
Adaptive.Systems.Theory
Steven E. Walls, Independent Consultant, USA
Abstract
This chapter seeks to identify the core of complex adaptive systems (CAS)
theory. To achieve this end, this chapter introduces innovative methods for
measuring and advancing the validity of a theory by understanding the
structure of theory. Two studies of CAS theory are presented that show how
the outer belt of atomistic and loosely connected concepts support the evo-
lution of a theory; while, in contrast, the robust core of theory, consisting
of co-causal propositions, supports the validity and testability of a theory.
Each may be seen as being derived from differing epistemologies. It is hoped
that the tools presented in this chapter will be used to support the purposeful
evolution of theory by improving the validity of intelligent complex adaptive
systems (ICAS) theory.
Walls
What is the Core of CAS Theory?
Where other chapters in this book may use intelligent complex adaptive
systems (ICAS) theory as a framework to understand our world, we strive in
this chapter to understand theory, itself. Through this process, the reader will
gain a new perspective on the theory that is applied elsewhere in this book.
To gain some perspective on ICAS, we will study the literature of complex
adaptive systems (CAS) as developed in the feld of organizational theory.
As such, this chapter may be of interest to those discussing organizational
theory and organizational change, multi-agent systems, learning methods,
simulation models, and evolutionary games.
CAS theory originated in the natural sciences as a tool for understanding non-
linear dynamics (Kauffman, 1995) and has gained popularity in organizational
studies through the efforts of many authors (i.e., Axelrod & Cohen, 2000;
Brown & Eisenhardt, 1998; Gleick, 1987; Stacey, 1996; Wheatley, 1992).
As CAS expanded into this discipline, every author seems to have placed a
personal mark by revising CAS for interpretation and publication. Indeed,
in researching the literature, 20 concise, yet different, defnition/descriptions
of CAS theory were found.
Within these 20 defnitions, component concepts were identifed. For ex-
ample, Bennet & Bennet (2004) note (in part) that a CAS is composed of a large
number of self-organizing components. The concepts of self-organization
and large number of components may be seen as conceptual components
of CAS theory as described by those authors. These conceptual components
might also be thought of as the authors propositions. It is important to
note that among the 20 defnitions, no two contained the same combination
of component concepts. This raises a serious question: When we talk about
CAS theory, are we really talking about the same thing? After all, if one
author states that a CAS may be understood through concepts a, b, and c
while another author states that the relevant concepts are c, e, and f, there
may be some conceptual overlap but there are also inherent contradictions.
In the social sciences, this issue has been of concern for decades. In one at-
tempt to make sense of the issue, theory has been described as consisting of
a hard core of unchanging assumptions, surrounded by a more changeable
protective belt (Lakatos, 1970). When a theory is challenged, a theorist may
rise to defend it with a new proposition that changes the belt, but presumably
leaves the core intact. Phelan (2001) suggests that complexity theory has its
hard-core assumptions; however, among the 20 defnitions discussed here,
there is no one concept that is held in common by all of the authors. If there
is no concept, or set of concepts, held in common, where then is the core of
CAS theory? Motivated by this apparent lack of commonality, we seek to
identify the core of CAS theory.
A Chinese proverb states, The beginning of wisdom is to call things by their
right names. (Unknown, 2006, p. 1). The diffculty of engaging in conversa-
tions with imprecise defnitions was famously illustrated when Plato called
man a featherless biped. He was forced to add, with broad fat nails in
response to Digenes, who arrived with a plucked bird, proclaiming Here is
Platos man. (Bartlett, 1992, p. 77). We may speculatively ask if it was the
atomistic nature of Platos defnition that left it so open to misinterpretation.
In short, we must wonder how we can know what we are talking about, if
the name keeps changing. In contrast to Platos rapidly evolving defnition,
Newtons laws (e.g., F=ma) have proved effective, and unchanged, for cen-
turies.
As scholars, of course, we are continually engaged in the discovery (or social
construction, depending on your view) of understandings and defnitions.
And yet, we need some level of shared understanding of existing concepts,
so we may communicate effectively as we work to understand new concepts.
In short, as scholars, we might see the increasing clarity and stability of our
defnitions as an indicator of progress in a given feld. We dig the clay,
form it into paving stones, place them in front of us, and walk on them to
fnd more clay. In this chapter, we will suggest some tools for identifying
the milestones along our shared road.
Central to the exploration presented in this chapter, we must ask, Is it pos-
sible to ascertain the legitimacy of a theory through its structure? Dubin
(1978) suggests that there are four levels of effcacy in theory; and these
levels do refect the structure of the theory. They are: (1) presence /absence
(what concepts are contained within a theory). (2) Directionality (what are the
causal concepts and what are the emergent concepts within the theory). (3)
Co-variation (how several concepts might impel change in one another). (4)
Rate of change (to what quantity do each of the elements within the theory
effect one another). Parson and Shills note four similar levels of systemization
of theorymoving toward increasing levels of systemization (Friedman,
2003, p. 518). Refecting the validity of these assertions, Newtons formulae
might be seen as residing at the highest level because it is possible to identify
quantitative changes in one aspect (e.g., force) from changes in other aspects
Walls
(e.g., mass & acceleration). Such a high level of understanding has been long
sought in the social sciences but has, as yet, remained elusive. One goal of
this chapter is to advance CAS theory along this scaleand identify how
further advances might be enabled for similar forms of theory.
To fnd the core of a theory, two studies are presented in this chapter. One
study is based on content analysis (essentially, looking at the words used
by the authors as reasonable representations of the concepts that they are
conveying). The second study uses a more traditional narrative analysis.
The frst study is a reductive look at CAS theoryfocusing on the axiomatic
propositions of the authors. This method will be seen as adding to the outer
belt. The second study focuses on the relational propositions of CAS theory.
This method suggests that there is a core to CAS theory. However, it also
shows that the core (based on the current state of CAS theory) has only a
limited internal integrity. A path for developing a more robust CAS theory
is then suggested. The process of developing a robust theory is expected to
provide great benefts to scholars (based on the successful use of Newtons
robust formulae).
Due to limitations of space, the studies in this chapter will be focused on the
level of concept (with concepts presented as they are named by the authors
and as they may be generally understood by most readers) and theory (as a
collection of concepts). These studies will generally avoid the sub-concept
level of interpretation and what might be called a post-theory level of ap-
plication and testing.
The next section includes a relatively linear and reductive analysis of the
concepts of CAS theory. This process might be seen as a thought experi-
menta cognitive construction that represents the creation of new defnitions
in an ad-hoc manner.
A.Reductive.Study.of.CAS.Theory
In this section, we engage in the development of new theory where theory
might be seen as a collection of concepts. This process identifes the range of
concepts in CAS theory and develops new versions of that theory. The new
versions of CAS theory created here may be seen as newly evolved defni-
tions. Although such defnitions may be tentatively used to identify various
perspectives of CAS theory, they may also be seen as adding to the outer
belt of theory rather than clarifying the core.
We begin with a review of literature. Searches of the ProQuest database yielded
nearly 100 articles in academic journals where CAS theory was discussed
in the context of a human organization. Within those articles, 13 were found
to contain concise (less than one page) defnitions of CASs. Additionally,
those journals (and other sources) suggested other scholarly publications.
Promising books were reviewed and seven additional concise defnitions
were found. In all, this study (although not exhaustive) found 20 relatively
concise defnitions of CASs. Concise defnitions were used so that the study
could cover as much ground as possible. It is also assumed that a concise
defnition includes the most important aspects of each authors version of the
theory. It is also expected that a sample of this size will provide a suffcient
representation of the body of theory.
Although the authors defnitions are not listed for reasons of space, this
study uses concise defnitions from Ashmos, Huonker, and McDaniel (1998),
Axelrod et al. (2000), Bennet et al. (2004), Brown et al. (1998), Chiva-Gomez
(2003), Daneke (1999), Dent (2003), Harder, Robertson, & Woodward (2004),
Hunt & Ropo (2003), Lichtenstein (2000), McDaniel, J ordan, & Fleeman,
(2003), McGrath (1997), McKelvey (2004), Moss (2001), Olson & Eoyang
(2001), Pascale (1999), Shakun (2001), Stacey (1996), Tower (2002), Yel-
lowthunder & Ward (2003).
In the process previously described, the study deconstructed each defnition
into the authors propositions, or component concepts. For example, Daneke
describes a CAS as, A simulation approach that studies the coevolution of a
set of interacting agents that have alternative sets of behavioral rules (Daneke,
1999, p. 223). The concepts describing the CAS here would be coevolution,
interaction, agents, and rules. While another reader might develop a different
list, it is expected that such lists would be substantively similar to the one
developed here where a total of 26 concepts were identifed, consisting of:
Agent, non-linear/unpredictable, levels, co-evolutionary, adaptive, agents
evolve, far from equilibrium/edge of chaos, self-organizing, many agents,
interrelated/interacting, goal seeking, decision-making, emergence/surprise
happens, act in rules/context of other agents and environment, simple rules,
permeable boundaries, evolves toward ftness, boundary testing, iterative
process, agents are semi-autonomous, evaluate effectiveness of decisions/
results,self-defning,identity,morality,irreversible,time.
Walls
This list might be seen as representing the whole of CAS theory from an
atomistic perspective. It should be noted that at this survey stage, no
component appears to be more important than another and no component
seems to be closer to the core than any other.
Of the 20 publications, three could clearly be seen as the most cited (each
having been cited by hundreds instead of tens, or fewer). Between them, there
are six concepts used by at least two of the three, including:
Co-evolutionary, many agents, interrelated/interacting, goal seeking, emer-
gence/surprise happens, simple rules.
This focus on what might be considered the authoritative versions essentially
creates a new defnition of CAS theory built on the shared conceptual com-
ponents of the authors. However, it should be noted that this new defnition
has lost some conceptual breadth when compared to the whole body of CAS
theory. Moving from one form of popularity to another, the following is a list
of those six concepts that seemed most popular among the 20 defnitions:
Non-linear/unpredictable, co-evolutionary, many agents, interrelated/inter-
acting, goal seeking, emergence surprise happens.
Again, a new defnition of CAS theory has been created with a new focus.
Again, the conceptual components have shiftedboth in comparison to the
whole body of CAS concepts and in comparison to the authoritative ver-
sion.
Additionally, while most authors identifed themselves as scholars, others
identifed themselves as scholar-practitioners. Those whose affliations were
uncertain were left out of this ad-hoc, demonstrative study. The fve concepts
most commonly described by those authors who identifed themselves as
scholars were:
Non-linear/unpredictable, self-organizing, many agents, interrelated/inter-
acting, emergence/surprise happens.
Among those who identifed themselves as scholar-practitioners, the four
most popular concepts used were:
Non-linear/unpredictable, many agents, interrelated/interacting, goal seek-
ing.
There are obvious limitations to this ad-hoc study. However, a number of
insights and benefts become apparent here. First, that this study creates a
comprehensive view of the concepts within a body of theory. Of course, in
this study, that view is limited to the level of the concepts, rather than delving
deeper, which is another possible level of exploration. Second, that each group
of concepts suggests those specifc concepts that might be most appropriate
for a given application or area of research. In a sense, each group of concepts
might be viewed as a school of thought for CAS theory within its specifc
venue. Importantly, this brief study essentially began with 20 defnitions
and generated four more. The number of theories in the outer belt was easily
increased, yet we do not seem to have increased our understanding of the
core. Our lack of core insight may be related to the form of analysis, or the
type of data used. Importantly, we approached the data as lists of concepts
and rearranged them into new lists. Each attempt to identify a new perspec-
tive resulted in a new list. As with the broader survey (the frst list), each
subsequent list has no discernible core.
The analysis presented in this section has served to demonstrate some strengths
and weaknesses of a reductive form of study. In the next section, we look
at alternative approaches to the ordering of conceptual components and, in
the section following, we apply those ordering ideas to clarify the structure
of CAS theory.

Looking.at.the.Structure.of.Theory
Drawing on Southerland, Weick (1989) discusses theory as, an ordered set
of assertions (p. 517). If a theory is defned (in part) as consisting of ordered
assertions, it begs the question of just how well ordered those assertions
might be. By ordered, we might understand those assertions to be arranged
Walls
alphabetically, by apparent importance, or any number of possible methods.
This disposition of things following one after another (Websters, 1989,
p. 1013) do not seem to add much to our understanding of theory, however.
It is not clear, for example, that a theory might be considered more valid if
the assertions are in alphabetical order instead of ordered (for example) by
the year each concept was added to the literature. However, based on that
simple interpretation of order, there would seem to be no epistemological
preference between ordering the assertions by their historical appearance in
the literature, by the frst letter in the concept, or ordering the assertions by
the apparent importance ascribed to them by an author.
By ordered, therefore, it may be that Weick was implying something more
signifcant than a list. A more useful (or at least an alternative) epistemo-
logical validity, therefore, might be developed by looking at the assertions
or propositions of a theory as being interrelated, where the propositions
might be seen as, reciprocally or mutually related (Websters, 1989, p.
744). With such a view, a body of theory might be seen as a kind of system
and, any part of the system can only be fully understood in terms of its
relationships with the other parts of the whole system. (Harder et al., 2004,
p. 83, drawing on Freeman). It seems, therefore, that every concept within
a theory would best be understood through other concepts within that body
of theory. Signifcantly, this perspective seems to ft with Dubins (above)
assertion that theories of higher effcacy have explanations and concepts that
are co-causal.
To briefy compare and contrast levels of interrelatedness, we might say that
the lowest level of relationship may be found in some jumble of random
concepts. A higher level of interrelatedness might be seen in a book where
an author describes concepts (thus causing each to exist in closer relation-
ship with others). Other authors have used a wide variety of methods for
increasing relatedness such as placing them in a list (as above), a fow-chart
showing a cycle (e.g., Nonaka, 2005, for social construction), a matrix (e.g.,
Pepper, 1961, for metaphors), or a combination of lists and fows to create a
meta-model (e.g., Slawski, 1990). With each increasing level of relatedness,
a given reader might understand a concept in relationship with other con-
cepts, and so fnd new insights based on the relatedness between concepts.
In short, an increasingly systematic relationship might be viewed as having
increasing relatedness. One example of a systemic theory may be seen in
Wilbers integral theory of human development (e.g., Wilber, 2001). In his
theory, Wilber describes four quadrants that represent categorizations of
insights from numerous disciplines. Wilber claims that each of these quad-
rants is co-defned by the othersin essence, that no quadrant can be fully
understood except in relation to the other three. This claim suggests a high
degree of relatedness.
Another way to look at interrelatedness might be seen in the concept of re-
fexive. Hall (1999) suggests that some forms of inquiry represent a third
path of inquiry that is primarily neither objective, nor subjective; rather it
is essentially refexive, where meaning is created in a socially constructed
sense. In contrast to refexive forms used in the sense of the interaction
between individuals, however, the second study of this chapter looks at re-
fexive analysis in the sense that suggests a relationship within, or between,
the concepts of CAS theory.
Combining the idea of relatedness with the idea of theory having a tight core
and a loose belt, we might see the concepts in the core as being more closely
interrelated than the belt. For example, the above reductive study of CAS
produced defnitions with low levels of interrelatedness, as might be found
in the loosely defned belt of a theory, because the new theories are presented
essentially as lists of concepts.
In addition to the concept of relatedness, another important concept for this
chapter is that of robustness. Wallis (2006a) explored a number of interpre-
tations of this term. Following insights developed from Hegel and Nietzsche,
Wallis settled on an understanding of robustness that might be familiar to
those working in the natural sciences, where a robust theory is one where its
dimensions are co-defned. An example of this would be Newtons law of
motion (F=ma) where each aspect (e.g., mass) may be calculated, or under-
stood, in terms of the other two (e.g., force and acceleration).
It is very important to differentiate between theories whose structure might
be seen as robust, and theories whose structure might be understood as an
ordered list. For example, a list of assertions might be understood as an
atomistic form of theory and might be represented abstractly as A is true.
B is true. C is true. In contrast, the propositions of a robust theory might
be seen as Changes in A and B will cause predictable changes in C.
Changes in B and C will cause predictable changes in A. And, changes
in C and A will cause predictable changes in B. The interrelatedness
of concepts in a robust theory suggests that the theory may be validated from
within the theory.
In this book and elsewhere (e.g., Richardson, 2006), the concept of robust-
ness is used to describe the stability of a network experiencing external per-
0 Walls
turbations. A system that is completely unstable would have a robustness of
zero, while a perfectly stable system would be assigned a robustness of one.
While it could be legitimately argued that no system can have its measure of
robustness at the extreme ends of the scale (zero, or one), this chapter will
use zero and one as approximations to facilitate discussion.
An understanding of perturbations might be used to determine what might
be called the dynamic robustness of a system of theory by identifying the
ratio of stable concepts to changing concepts. In this two-step process, one
frst identifes the concepts contained in each form of the theory and assigns
each a numerical value based on the component concepts. Next, the ratio
between the two (earlier and later) versions of the theory is taken. If the two
theories are identical, the robustness will be equal to unity (or one). If the
two theories have no concepts in common, they will have a robustness of
zero. For example, if theory A has four distinct concepts (a, b, c, d) and
subsequently evolves into theory B with four concepts (c, d, e, f,), we may
see theory A and B together as having a total of six concepts (a, b, c, d, e, f)
with only two concepts held in common (c, d). This relationship suggests that
in the process of evolving from theory A to theory B, the theory exhibited
a robustness of 0.33 (two divided by six). Of course, such measures might
only be considered valid when the concepts themselves are unambiguous.
This method may be seen as responding to Hulls (1988) deep discussion
on the evolution of theoryand providing a tool to aid in the mapping of
that evolution.
If we look at each authors infuence on CAS theory as a perturbation, CAS
theory may be seen as having a low level of robustness. For example, Yel-
lowthunder et al. (2003) describe a CAS using four concepts drawn from
Olson et al. (2001) who used those four in addition to three additional con-
cepts. This change suggests a robustness of 0.57 (the result of four divided
by seven). Other times, CAS does not fare even that well. Dent (2003) states
that he drew his conceptualization of CAS from Brown et al. (1998). How-
ever, between the six concepts listed by Dent and the eight concepts listed by
Brown et al., there are only two concepts that clearly overlap. This suggests
a low robustness of 0.14 (the result of two divided by fourteen). In contrast,
Newtons formula of force (F=ma) may be seen as having a robustness of
one as the formula is unchanged in any non-relativistic application. The
widespread use of Newtons formula may suggest that theories of greater
robustness are more useful (and may have more predictive power) than
theories of less robustness.
While low robustness may be seen as enabling fexibility (where a theory
changes and evolves with rapidity), it may also be seen as an indicator of
confusion or uncertainty. It recalls our original question as to the core of
CAS theory and suggests that axiomatic, atomistic, or reductive defnitions
(e.g., theories with concepts that are structured as lists) have shown too much
fexibility to provide an adequate representation of the core. In the following
section, we analyze the body of CAS theory to identify more robust relation-
ships between the concepts.
Investigating.Relational.Propositions
In this section, we investigate the relational propositions described by the
authors of the above 20 concise defnitions to identify the core of CAS theory,
where that core may be seen as shifting CAS theory toward Dubins (1978)
second level of theory effcacy (where concepts are directionally causal).
In this process, as with the reductive study, we deconstructed each of the
concise defnitions into propositions. Rather than use all of the available
conceptual components, however, those statements that were essentially
axiomatic are left out. For example, Bennet et al. (2004) state, There are
some basic properties common to many complex adaptive systems. Examples
are some level of self-organization, nonlinearity, aggregation, diversity, and
fow (p. 26). Those concepts would be considered axiomatic or atomistic.
In contrast, their statement that, The term complex system means a system
that consists of many interrelated elements with nonlinear relationships that
make them very diffcult to understand and predict (p. 26) may be seen as
a relational proposition because nonlinear relationships are seen to cause
unpredictability.
Of the few relational propositions, the frst is Stacey (1996) who states that
agents follow rules (or schemas) in their interactions to improve on their
behavior. This proposition shows that there is some relationship between
the agents, their schemas, behaviors, and the subsequent improvement in
behavior. Many authors echo this same general idea.
Axelrod et al. (2000) note that varied schemas (situational decision-mak-
ing rules) differentiate, or provide variety, among agents. Agents are also
differentiated by geography (differences in physical and conceptual space).
These agents interact with one another (and with tools) in an essentially evo-
Walls
lutionary process that might be seen as being based on the agents ft with the
environment. In that process, the agents are changed through changes in their
schemas. Changes might be seen as increasing or decreasing the similarity of
those agents. Similarly, Shakun (2001) states that agents take actions to reach
goals. In this conversation, goals may be seen as generally similar to schemas
as both seem to have some infuence over the actions of the agents.
McDaniel et al. (2003) also suggest that agents interacting over time leads
to self-organization. Time may be seen as important, although most authors
include it only implicitly. Moss (2001) notes that members (agents) self-or-
ganize toward more stable patterns of activity. This may be seen as generally
similar to the process of agents interacting to cause self-organizationswith
the added idea that the process of self-organizing causes more stable patterns
of activity. These stable patterns of activity are a result of common frames of
reference, an idea that seems generally similar to rules or schemas as noted
above.
Other authors (e.g. Bennet et al., 2004; Hunt et al., 2003) start from generally
the same positionthat the components (which might be seen as agents) of
a CAS interact. However, according to these authors, the interactions lead to
uncertainty. Dent (2003) also agrees and adds that the interaction is to fnd ft.
Dent then notes the results of the agentic interactions may be seen as causing
change (that may be thought of as a form of uncertainty).
Somewhere between, or combining, these camps, Harder et al. (2004) state
that varied agents interact to maintain a system (homeostasis, in their words)
rather than create a new system. However, their description of homeostasis
is described by terms such as dynamic equilibrium, constant change,
and adaptation. Thus, it seems that the authors are suggesting that agents
do not changeso much as they enable their CASs to interact, change, and
evolve.
Pascale (1999) notes that the process of agents engaged in interaction will
lead to more levels of organization. This idea of levels might be seen
as conceptually similar to Axelrod et al.s description of the geographic
differentiation of agents (physically and/or conceptually). Additionally, it
does not seem as though the creation of a new level of organization should
be signifcantly different (within the context of this conversation) than the
creation of a new system. It simply seems that this particular new orga-
nization is one that is already nested in an existing one. A larger difference
between Pascales version and that of other authors is that Pascale states that
the agents are shuffed by the larger system.
Drawing on Dooley, Olsen & Eoyang (2001) note that agents evolve over time
to reach ftness. Again, we may note the explicit surfacing of the temporal
aspect of CASs that many authors leave tacit. Also, where most relational
statements discuss agents interacting to achieve ftness, these authors might
be seen as leaving the step of interaction as tacit. In both versions, however,
agents do tend toward ftness.
Finally, Chiva-Gomez (2003) draws on Stacey to note that CASs that are
closer to the edge of chaos (EOC) will experience more self-organization.
This concept of EOC might be considered synonymous with bounded in-
stability, which Stacey describes as a balance between formal and informal
systems. More broadly, EOC might also be described as the boundary be-
tween stability and ambiguity. In developing the OEC concept, Stacey (1996)
notes Kauffman, Wolfram, Gell-Mann, and Langton among his infuences.
Seen from the perspective of the present conversation, Kauffmans (1995)
description of bounded instability may be understood as occurring where the
number of agents approximates the number of interactions. If there are too
many agents, and insuffcient interactions, chaos reigns. On the other hand,
if there are few agents and many interactions, stability prevails. Therefore,
it seems reasonable that we may integrate the EOC concept with the other
co-causal statements above because the EOC may be understood as a ratio
between the number of agents and their interactions.
Generally speaking, there appears to be considerable overlap between many
of the previous causal propositions. Specifcally, many authors discuss the
existence of agents (including parts), schemas (that may be seen as including
goals, rules of interactions), interactions (that may include communication, and
also implicitly or explicitly assumes the passage of time), and ft (including
evolutionary tests of success). Additionally, it may be seen that the ft test is
based on the existence of an external environment (although that environment
may be seen as one or more other agents). There are a variety of changes that
may result from a ft test including adaptation, change in interaction, change
in schemas, increasing uncertainty, increasing certainty, self-organization,
and the maintenance of existing organization. Also, as change is seen as a
general result, change may also be seen to alter agents.
Looking at the concepts in their causal relationship, we may defne the core
of CAS theory as agents, with schemas, interacting over time. The results
of those interactions are maximized at the EOC and are subject to a ft test
with the environment. The result may be changes in schemas, changes in
interactions, the creation or maintenance of larger systems, increased stability
Walls
and increased instability. Finally, the status of the EOC may be changed by
the creation of new agents, schemas, or CASs. This defnition is represented
graphically in Figure 1.
Each arrow represents causal direction, where one aspect of CAS theory
will have an effect on another. As the concept of non-linear dynamics is an
important aspect of CAS theory and complexity theory (e.g., Dent & Holt,
2001; Lichtenstein, 2000), it may be worth noting that the relationship between
the causal aspects of this defnition of CAS theory seems to support the idea
of a non-linear or non-deterministic structure. This view is in contrast to an
atomistic list of conceptual components.
The presented defnition surfaces additional challenging questions. These
questions might be seen as stemming from potentially contradictory state-
ments by the authors. For example, Axelrod et al. note that changes occur at
Figure 1. Relationship between causal concepts
the level of the schema while Moss suggest that change occurs at the level
of interactions. These should not, however, be seen as mutually exclusive.
Instead, we might ask, How much change will occur? Then we might ask,
Where will that change occur? For example, if a group of individuals self-
organize into a corporation to take advantage of a new business opportunity,
how much change is seen from the perspective of the business environment
(with the inception of a new entity) compared to how much change is seen by
the agents (as they organize themselves into new relationships and interac-
tions), and how much change might be seen as occurring at the level of the
schema of those individuals?
The structure of this model does not seem to be perfectly causal, however,
in at least three areas. First, time (along with agents and change) enable the
occurrence of interactions. However, nothing seems to cause (or alter)
time. Therefore, time seems to be seen as an atomistic concept. To become
more fully robust, this model should identify what causes (and alters) time
or/and its perceived passage. Second, schemas seem to be caused by change,
alone. Such a linear relationship would seem to suggest that any change in
schemas would create a corresponding change in agents. This relationship
would suggest that the concept of schemas might be bypassed. Alternatively,
the concept of schemas might be enhanced by adding causal infuences. Fi-
nally, while the EOC seems to control the quantity of change, the variety
of possible forms of change is still open to interpretation. Similarly, the
environment seems to be changed only by change (another linear relation-
ship, like schemas). However, the environment may represent numerous
CASseach with their own nested agents and schemas.
Although this model has left out many component concepts of CAS theory,
the closely related nature of the included aspects suggest that it may be a
good representative for the core of CAS theory. That is because that causal
model more closely meets Dubins requirement for stage two of effcacy of
theory, while the reductive models only reach stage one. In the next section,
we compare three models of CAS theory.
Comparisons.and.Insights
Drawing on Davidson and Layder, Romm (2001) suggests that research-
ers use triangulation (multiple research methods to reduce subjectivity
Walls
in research). Thus far, we have presented two forms of analysis. A third
form of analysis of the same data set may be found in Wallis (2006a) where
refexive dimensional analysis (RDA) was used to understand CAS theory
as consisting of fve related conceptual dimensions. Developed by Wallis
(2006b) from grounded theory and dimensional analysis, RDA is a method
for the investigation of a body of theory.
Looking at the broad range of concepts in CAS theory, the previous reduc-
tive study has the beneft of including all concepts of the body of CAS
theorypresented as a list of atomistic concepts. The RDA version also
purports to include all the concepts of CAS theory; however, it presents those
concepts as enfolded so that concepts not directly represented by the theory
might be understood indirectly by combining the dimensions of the theory.
For example, the concept of evolution (not directly represented) might be
understood as change over time (both of which are directly represented
in the model). In contrast to either of the above methods, the causal model
leaves out concepts that are considered to be axiomatically atomistic (although
those concepts may be seen as still existing in the belt). This suggests some
benefts and detriments to each method. Where the atomistic/reductive ap-
proach might be more complete, the RDA version might be understood as
more abstract (and so has the opportunity to be applied to a wider variety of
systems) and more parsimonious (where aspects are related by the minimum
number of laws possible, per Dubin).
Comparing the fexibility of these forms of theory (setting aside their common
requirement for scholarly justifcation), it seems that reductive forms enable
the easy addition or removal of concepts, essentially adding or removing them
from the list. In contrast, forms of theory that are refexively structured (RDA
and causal) are not so easily altered. For example, referring to Figure 1, if
we were to remove the concept of interactions, our understanding of other
concepts would be imperiled. Time would become disconnected from the
model, agents would become linearly causal to change (thus eliminating the
concept of EOC), and agents would go straight to ft test without interactions
(a wholly unsatisfactory description).
In short, it seems that the reductive (and less integrated) forms of theory are
more easily changed and therefore may be seen as more easily manipulated
or evolved by theorists. That, in turn, suggests that the closer we get to a
fully robust form of theory, the more diffcult it will become to make further
progress. The next few decades, therefore, may see the evolution of compet-
ing very-nearly robust theories. Those theories may then be tested (in real
world and computer modeling venues). Then too, each version of theory may
fnd its own specifc niche (e.g., one model may be applied to organizations,
another to individuals, and a third to schemas).
Additionally, the reductive study suggests a technique for identifying which
concepts within a general body of theory might be more closely connected
with a given focus (e.g., scholarly version of CAS). The RDA and causal
versions, in contrast, are both highly integrated so that the researcher is
encouraged (almost required) to utilize all of the concepts for any given
analysis. In the causal model depicted in Figure 1, for example, a researcher
focusing on interactions would be impelled to describe how those interactions
are changed by changes in agents, change, and time. Additionally, the causal
relationships would suggest that the researcher describe how changes in the
interactions aspect altered the EOC and the ft test. In a sense, this creates a
road map that might be of beneft to researchers and students alike.
Shifting our focus to the core, we should note that the core concepts of the
RDA version of CAS are (atomistically) interactions, agentness, change, levels
of difference, and time. The causal version of CAS may be represented as
schemas, agents, interactions, time, ft test, change, EOC, and environment.
The two models hold in common the concepts of change, interactions, and
time. In the RDA version, the schemas, agents, and environment of the causal
version might be explained as agentness seen at different levels. Both of these
models suffer from understanding time as an atomistic concept. Where time
enables change (for example), nothing enables time. This could seem to be
a relatively innocuous concept; however, the concept of fow (Csikszent-
mihalyi, 1991) suggests that the idea of time, especially as a subjective rep-
resentation of productivity, may be an important area for investigation. The
RDA version leaves the idea of agents (and other forms of systems) loosely
defned based on the idea that it is the observer who determines what the
system is. The causal version leaves the concept of Levels relatively tacit,
seemingly accepting that there are three levels of systems (which may be
broadly understood as schemas, agents, and environment). Similarly, there are
differences within each causal concept (there are different forms of change,
a variety of agents, and schemas include a variety of rules).
In contrast to what might be called the dynamic measure of robustness
discussed previously in this chapter (where we quantifed the level of robust-
ness of a theory as it evolved between authors), we might apply here a more
static measure of robustness. Static robustness might be applied to quan-
tify the internal integrity of the structure of a body of theory and so provide
Walls
a point of comparison between the two views of the core. Alternatively, it
may be used to differentiate between core and belt concepts within a body
of theory. A measure of robustness may be achieved by comparing the total
number of aspects within the theory to the number of aspects that are both
causal and emergent. On one end of that scale, a theory that simply lists its
component concepts, without identifying how the concepts were related to
one another would have a robustness of zero. On the other end of the scale, a
fully robust theory (e.g., Newtons F=ma) might have three dimensions, each
of which is both causal and emergent, and therefore would have a robustness
of one (3 / 3 =1). By emergent, it is important to note that a given concept
must be understood in relation to two or more other concepts, as in Newtons
model. If changes in one concept were to be determined directly by only one
other concept, that change would be seen as linear, and so adding little to
the model. For example, if we were to say that changes in the environment
cause changes in the agent that cause changes in the schema, we might as
well say that changes in the environment cause changes in the schema. If, on
the other hand, there is something about the agent that ameliorates, flters,
or accentuates, the changes from the environment, we may then say that the
changes in the environment and the effect of the agent result in changes to
the schema.
In the previous reductive study, all lists have robustness of zero (e.g., 0 / 29
= 0). In the RDA CAS model, there are a total of fve dimensions (agentness,
levels, change, time, and interactions). However, of those fve, only three are
both emergent and causal (agentness, interactions, and levels). Therefore, we
might understand this form of the theory as having a static robustness of 0.6
(3 / 5 =0.6). In contrast, the causal version presented in this chapter contains
eight concepts only fve of which may be seen as both causal and emergent
providing a robustness of 0.625 (5 / 8 =0.625). In short, the causal version
may be seen as an improvement over the RDA version and this chapter serves
as an example of theory-advancement towards a robust core.
Shifting to the relatively non-robust aspects, we see that time appears in
the causal model to be atomistic. Schemas and environment are seen as
linear/determinant. In the causal model, it seems that the more important
area for investigation are the concepts of time, environment, and schema.
Understanding how those concepts may be understood as emerging from
two or more other concepts within the core should indicate how the model
might be rearranged to make sense from a robust perspective. For example,
the multiple levels of this model (including schema, agents, and environ-
ment) may each be seen, in some sense, as a CAS. That may indicate the
opportunity to create a model where each level is represented by that same
(simpler) model. Such repetition of simpler models might, in turn, allow
the elimination of some redundancies from the model. Indeed, until such an
investigation is undertaken, those linear components of an otherwise robust
model might be seen as existing somewhere between the belt and the core
in an intermediate, or connecting zone.
Finally, based on Dubins list (where increasing relatedness suggests higher
effcacy of theory) and Weicks inference (that propositions should be ef-
fectively ordered, or interrelated), it may be suggested that increasing ro-
bustness suggests a higher level of epistemological validity of a theory. In
general, however, it is important that the causal core may be seen as being
derived from a different epistemological validity than the loosely related
list of concepts that comprises the belt. Where the belt may fnd validation
from any one of a wide range of research methodologies and points of view,
the core fnds validation only in the relationship between its own concepts.
This is a signifcant epistemological shift that suggests a rich opportunity
for additional study.
Will CAS and/or ICAS Theory Survive?
In this chapter, 20 concise versions of CAS theory were found in the disci-
pline of organizational theory. The conceptual components were identifed
for each theory and subjected to two forms of analysis. The frst, a reductive
study, was benefcial in the identifcation of concepts representing the range
of CAS theory, for linking specifc concepts within the body to specifc uses
of the theory, and the creation of additional versions of CAS theory. Each
additional defnition was seen as adding to the outer belt of the theory rather
than clarifying the core. The second study focused on the causal statements
found within the body of CAS theory and identifed the core of CAS theory
by identifying relationships between those concepts. It is suggested that by
focusing on causal relationships, we may be able to accelerate the evolution
of CAS theory.
0 Walls
This chapter suggests that the fexibility of loosely connected forms of theory
may support the spread of that theory through the social sciences. However,
the very fexibility that allows CAS theory to grow may also obscure the
core. The lack of core, in turn, may limit the effectiveness of that theory in
application. Another signifcant contribution of this chapter is the creation of
specifc measures of robustness as tools for examining theory. These include
measures of dynamic robustness (as an indicator of the evolution and stability
of a theory as it passes between authors), and static robustness (a numerical
indicator of how well co-defned a theory may be). The more static robustness
a theory has, the more it may be considered to be part of the core.
Will CAS theory survive? Or, will it join the 90% of social theory that rises
rapidly only to disappear just as quickly (Oberschall, 2000)? This chapter
represents a signifcant step toward a new understanding of theory (in general)
and CAS theory (in particular). If CAS theory is to retain its validity and even
gain credibility in the face of the next wave of theory (that will, undoubtedly,
arrive), it seems that it must develop a robust core. In an important sense, a
robust theory might be seen as possessing epistemological justifcation, not
from external testing, but internally, as the understanding of each aspect of
the core is tested against the other aspects of the model. Conversations on the
structure and construction of theory are likely to continue, and even increase,
as our understanding of theory-creation increases. Measures of theory-robust-
ness would seem to provide useful tools for advancing that conversation.
Looking to achieve a more optimal future of CAS theory, investigations
should frst clarify the causal relationships suggested in this chapter. With
a fully robust version of CAS, the next step should be to test that model in
the feld and through computer modeling to clarify (or deny) relationships
suggested by the co-defned aspects. Additionally, the inter-testability of
the core aspects hints that such a model might be falsifable in the Popperian
sense. For example, if the model pictured in Figure 1 were to experience a
change in interactions that did not result in a change in the ft test, the model
would be disproved. This would, of course, open the door to improving the
model (Popper, 2002). Then too, if everything observable may be explained
in terms of the robust aspects, then each application in the feld becomes a
test of the theory, and we have another opportunity to accelerate the evolu-
tion of CAS theory through practice.
This chapter has focused on CAS theory, but what about its close cousin ICAS?
While this book may, or may not, contain all concepts related to ICAS theory,
it certainly provides a cross-section of that theory. As such, the reader may
apply the insights and techniques presented in this chapter as he or she reads
other chapters. That way, the reader may identify the full breadth of concepts
for ICAS theory. Similarly, within each chapter, the reader may fnd an em-
phasis on those concepts specifcally related to that particular area of study
and so suggest a school of thought within ICAS. Finally, the reader may
range between the chapters seeking causal relationships between concepts,
and so develop a robust model of ICAS theory. Each of these opportunities
suggests how the reader might develop an alternative point of view that may
be useful for further study and developing new insights.
In conclusion, we have presented three major insights that may support
theory development and the progress of ICAS theory. First, a robust form of
theory provides the best description of a solid core of a theory and so avoids
the growing belt of loose concepts that obscures it. Second, it is possible
and desirable to measure a theorys level of robustness and by so doing, to
measure the progress of that theory. A corollary here is that measuring the
progress of a theory opens the door to advancing that theory in a more rapid
and purposeful way. Finally (although less deeply explored), robust theories
may provide a path to more effective analysis and application. As the data
for developing the core came from the belt, it should be noted that no core
is possible without a belt. This suggests that both belt and core, with their
separate epistemological justifcations, are necessary to the advancement of
theory.
Shifting to an evolutionary perspective, CAS might be seen as a recently
evolved (and rapidly evolving) species of theoryderived from its fecund
progenitors in systems theory and complexity theory. As a relatively recent
species, it is well adapted to ft its niche as an insight-generator for theorists.
Theorists, theory, and this book (as a representative of the conversation)
may then be understood as three speciesall engaged in a co-evolutionary
process. This co-evolutionary process, in turn, suggests the opportunity to
improve ourselves by accelerating the evolution of CAS. In the sense of an
evolutionary landscape, this book might be seen as a path leading CAS off
of the plains (inhabited by herds of big-belt theories) and up the slope of
Mt. Kilimanjaro.
Walls
References
Ashmos, D. P., Huonker, J . W., & Reuben, R., & McDaniel, J . (1998). Partici-
pation as a complicating mechanism: The effect of clinical professional
and middle manager participation on hospital performance. Health Care
ManagementReview,23(4), 7-20.
Axelrod, R., & Cohen, M. D. (2000). Harnessing complexity: Organizational
implicationsofascientifcfrontier. New York: Basic Books.
Bartlett, J . (1992). Familiar quotations: A collection of passages, phrases,
and proverbs traced to their sources in ancient and modern literature
(16
th
ed.). Toronto: Little, Brown.
Bennet, A., & Bennet, D. (2004). Organizational survival in the new world:
The intelligent complex adaptive system. Burlington, MA: Elsevier.
Brown, S. L., & Eisenhardt, K. M. (1998). Competing on the edge: Strategy
as structured chaos. Boston: Harvard Business School Press.
Chiva-Gomez, R. (2003). The facilitating factors for organizational learning:
Bringing ideas from complex adaptive systems. Knowledge and Process
Management,10(2), 99-114.
Csikszentmihalyi, M. (1991). Flow: The psychology of optimal experience.
New York: Harper Perennial.
Daneke, G., A. (1999). Systemic choices: Nonlinear dynamics and practical
management. Ann Arbor, MI: The University of Michigan Press.
Dent, E. B. (2003). The complexity science organizational development
practitioner. OrganizationDevelopmentJournal,21(2), 82.
Dent, E. B., & Holt, C. G. (2001). CAS in war, bureaucratic machine in peace:
The U.S. Air Force example. Emergence,3(3), 90-107.
Dubin, R. (1978). Theory building (Revised Edition). New York: The Free
Press.
Friedman, K. (2003). Theory construction in design research: Criteria: Ap-
proaches, and methods. DesignStudies,24(6), 507-522.
Gleick, J . (1987). Chaos: Making a new science. New York: Penguin
Books.
Hall, J . R. (1999). Cultures of inquiry: From epistemology to discourse in
sociohistorical research. New York: Cambridge University Press.
Harder, J ., Robertson, P. J ., & Woodward, H. (2004). The spirit of the new
workplace: Breathing life into organizations. Organization Development
Journal,22(2), 79-103.
Hull, D. L. (1988). Science as a process: An evolutionary account of the
social and conceptual development of science. Chicago: University of
Chicago Press.
Hunt, J . G. G., & Ropo, A. (2003). Longitudinal organizational research and
the third scientifc discipline. Group & Organizational Management,
28(3), 315-340.
Kauffman, S. (1995). At home in the universe: The search for laws of self-
organization and complexity (Paperback Edition). New York: Oxford
University Press.
Lakatos, I. (1970). Falsifcation and the methodology of scientifc research
programmes. In I. Lakatos & A. Musgrave (Eds.), Criticism and the
growth of knowledge (pp. 91-195), Cambridge University Press.
Lichtenstein, B. M. B. (2000). Emergence as a process of self-organizing: New
assumptions and insights from the study of non-linear dynamic systems.
JournalofOrganizationalChangeManagement,13(6), 526-544.
McDaniel, R. R., J ordan, M. E., & Fleeman, B. F. (2003). Surprise, surprise,
surprise! A complexity science view of the unexpected. Health Care
ManagementReview,28(3), 266-277.
McGrath, J. E. (1997). Small group research, that once and future feld: An
interpretation of the past with an eye to the future. Group dynamics:
Theory, research, and practice, 1(1), 7-27.
McKelvey, B. (2004). Toward a 0th law of thermodynamics: Order-creation
complexity dynamics from physics and biology to bioeconomics. Journal
of Bioeconomics, 6(1), 65-96.
Moss, M. (2001). Sensemaking, complexity, and organizational knowledge.
Knowledge and Process Management, 217.
Nonaka, I. (2005). Managing organizational knowledge: Theoretical and
methodological foundations. In K. G. Smith & M. A. Hitt (Eds.), Great
minds in management: The process of theory development (pp. 373-393).
New York: Oxford University Press.
Oberschall, A. (2000). Oberschall reviews theory and progress in social
science by J ames B. Rule. SocialForces,78(3), 1188-1191.
Walls
Olson, E. E., & Eoyang, G. H. (2001). Facilitating organizational change: Les-
sons from complexity science. San Francisco: J ossey-Bass/Pfeiffer.
Pascale, R. T. (1999). Surfng the edge of chaos. Sloan Management Review,
40(3), 83.
Pepper, S. C. (1961). World hypothesis: A study in evidence. Berkeley, CA:
University of California Press.
Phelan, S. E. (2001). What is complexity science, really? Emergence,3(1),
120-136.
Popper, K. (2002). Thelogicofscientifcdiscovery (J . F. Karl Popper, Lan
Freed, Trans.). New York: Routledge Classics.
Richardson, K. A. (2006, J une). The role of information barriers in
complex dynamical systems behavior. Paper presented at the NESCI,
Boston, MA, USA.
Romm, N. R. A. (2001). Accountability in social research: Issues and debates.
New York: Kluwer Academic/Plenum Publishers.
Shakun, M. F. (2001). Unbounded rationality. Group Decision and Negotia-
tion,10(2), 97-118.
Slawski, C. (1990, J uly 8-13). A small group process theory. Paper presented
at the Toward a J ust Society for Future Generations34th Annual
Meeting of the International Society for Systems Sciences, Portland,
Oregon, USA.
Stacey, R. D. (1996). Complexity and creativity in organizations. San Fran-
cisco: Berrett-Koehler Publishers.
Tower, D. (2002). Creating the complex adaptive organization: A primer on
complex adaptive systems. ODPractitioner,34(3).
Unknown. (2006, 3/6/07). Quotation. Retrieved March 3, 2007, 2007, from
http://www.quotationsbook.com/quote/27796/
Wallis, S. E. (2006a, J uly 13, 2006). A sideways look at systems: Identifying
sub-systemic dimensions as a technique for avoiding an hierarchical
perspective. Paper presented at the International Society for the Systems
Sciences, Rohnert Park, California.
Wallis, S. E. (2006b). A study of complex adaptive systems as defned by
organizational scholar-practitioners. Unpublished Theoretical Dis-
sertation, Fielding Graduate University, Santa Barbara.
Websters. (1989). Encyclopedic Unabridged Dictionary of the English
Language. Avenel, NJ : Gramercy Books.
Weick, K. E. (1989). Theory construction as disciplined imagination. Acad-
emyofManagementReview,14(4), 516-531.
Wheatley, M. J . (1992). Leadership and the new science. San Francisco:
Barrett-Koehler.
Wilber, K. (2001). A theory of everything: An integral vision for business,
politics, science, and spirituality. Boston, MA: Shambhala.
Yellowthunder, L., & Ward, V. (2003). Designing and implementing organi-
zational change in a complex adaptive system. In G. H. Eoyang (Ed.),
Voicesfromthefeld:Anintroductiontohumansystemsdynamics (pp.
125-144). Circle Pines, MN: Human Systems Dynamics Institute.
Pushno & Bonser
Chapter.II
Method.of.Systems.
Potential.as.Top-Bottom.
Technique.of.the.
Complex.Adaptive.Systems.
Modelling
Grgor S. Pushno, St. Petersburg, Russa
Gordon L. Bonser, Calforna, USA
Abstract
Emergent properties of complex adaptive systems (CAS) are explored by
means of agent-based modelling (ABM), which are compared with results
from modelling on the basis of the method of systems potential (MSP). MSP
describes CAS as a holistic system whereas ABM-methodology considers
CAS as set of interacting agents. It is argued that MSP is a top-bottom
approach, which supplements ABM bottom-up modeling of CAS. Adap-
Method of Systems Potental as Top-Bottom Technque of the CAS Modellng
tive principles incorporated into CAS at the level of a holistic system exploit
Lamarcks ideas about evolution, while the adaptive rules incorporated in
theinnerstructureofCASrefectDarwinsideas.BothABMandMSPex-
hibit the same macroscopic properties: (1) punctuatedequilibrium;(2)
sudden jumps in macro-indices;(3)cyclicaldynamics;(4)superpositionof
deterministic and stochastic patterns in dynamics; (5) fractal properties of
structure and dynamics; (6) SOC-phenomenon. ABM demonstrates these
properties via simulations of the different models whereas MSP derives these
phenomena analytically.
Introduction
Traditionally, in modeling the complex adaptive systems (CAS), multi-
agent modelling approach (MAM) is used. The complex adaptive system is
modelled as a multitude of agents interacting with each other in line with
certain rules of adaptive behaviour. Setting the rules of the agents adap-
tive behavior, the researcher has a possibility to trace, using the computer,
the behaviour of the system with time. Lately tremendous experience has
been gained in the study of macroscopic properties of such systems. As far
as these properties cannot be derived from the agents properties, they are
often called the emergent properties of the system. It has been found that
dissimilar MAM-models show the same set of macroscopic emergent proper-
ties when being modeled on the computer.
Universal.Emergent.Properties:
Punctuated equilibrium.
Self-organised criticality.
Superposition of deterministic and stochastic patterns in macroscopic
dynamics.
Discontinuous cycles.
Catastrophic jumps.
Self-affne dynamics of macro-indices.
Power law for avalanche-size distribution.
Perpetual renewal of confguration.
Pushno & Bonser
Creation of hierarchical fractal-like structure in course of evolution.
Episodic sudden reconfgurations of CAS-structure.
(*)
It is notable that the closer the model is to its real prototype, the more pre-
cise the intelligent agents reproduce the behaviour of real CAS agents,
the clearer and more defnitely these general macroscopic properties of the
models are displayed. This appears to imply that the emergent properties (*)
of the system do not depend on its internal structure or on specifc features
of the agents adaptive behaviour in a specifc model, but rather are in-
herent in CAS as a holistic system. This chapter shall attempt to prove that
the emergent properties of CAS may be explained proceeding from general
regularities observed in the evolutionary process.
Multi-agent modelling reveals the emergent properties by means of simula-
tion. Within the framework of -platform these properties emerge as
unpredictable aggregated macroscopic effects of interaction of a multitude
of agents. Multi-agent modelling is a method of studying CAS on a bot-
tom-up basis, starting from the interaction of a multitude of agents to
revealing the emergent properties of the integral system.
The departure point of our analysis of the integral complex adaptive system
is an ensemble of interacting agents. The nature of such systems supposes
the existence of some universal adaptive mechanisms effective at the level
of the integral system. The existence of simple and universal mechanisms
of adaptation acting at the level of integral systems has been a subject of
discourse with many authors.
Gell-Mann (1994, p. 17), Holland (1995), Arthur, Durlanf, and Lane (1997,
pp. 2-3), and McMillan (2004, p. 103) all note that the system is adaptive
if it (1) is permanently renewed at all levels, (2) organises a multi-level hi-
erarchical structure within itself, (3) draws information from the surround-
ing environment, (4) uses this information in its adaptive activity, and (5)
adjusts itself proceeding from the effciency of resultant gain. One may say
that CAS accumulates the useful adaptive experience through the process of
adaptive activity and applies the gained experience as a means of continued
survival.
It is assumed for this analysis that some universal rules of evolution can be
formulated in respect to integral systems. These rules describe the aggregate
macroscopic effect of agents interactions within the system under consid-
eration. This approach will be called the method. of. systems. potential
(MSP).
These adaptive mechanisms are effective at the level of the integral system.
They may therefore be called macroscopic mechanisms of adaptation. They
are working in any CAS. Therefore they may be called universal mechanisms
of adaptation. As far as any CAS comprises a multitude of smaller-scale
complex adaptive systems (modules or agents), the universal macroscopic
adaptive mechanisms (UMAM) function in each of these CA-sub-systems.
Thus, UMAM function at different levels of CAS, within each of its sub-
systems.
The effect of UMAM within CAS may be formalised within the framework
of some logical and mathematical scheme. The key notions are:
1. Adaptive potential of the system: The aggregated ability of CAS to
adequately respond to the challenges of the external world,
2. Conditions for realisation of adaptive potential: The aggregate fac-
tors contributing to (or preventing) the exploitation (employment) of
the adaptive potential,
3. Effciency of CAS: The relationship between the exploited adaptive
potential and the accumulated one.
Adaptive potential and conditions of realisation are fundamental terms
of MSP. These values describe the global steady state of CAS as ensembles
of interacting agents. They play the role of thermodynamic potentials for
these ensembles. Adaptive potential and conditions for its realisation are
two constituents of useful experience accumulated in a CAS.
If the adaptive abilities of an agent are greater, the ftness of this agent is
greater. Adaptive potential of CAS as a whole is the measurement of ag-
gregated adaptive abilities of whole agents within the ensemble. The ftness
of each agent is dependent upon both features of agent itself and from inter-
agent interactions. Adaptive potential is accumulating within the operative
memory and functions of agents via learning and training. Mutations fx
accumulated potential as information coded in the genetic structure.
So long as any CAS consists of complex adaptive sub-systems, the defnition
of adaptive potential of comprehensive CAS depends upon the defnition
0 Pushno & Bonser
of the adaptive potential of sub-systems (modules or agents). The lowest
level of real CAS structure is unknown, but we can defne adaptive potential
of CAS under consideration as the sum of adaptive potentials of its sub-
systems, additionally taking into account inter-sub-systems interaction.
The term conditions of realisation is the second thermodynamic potential
of ensembles of interacting agents. Each agent is limited in its choice of adap-
tive behaviour. Interaction with other agents and the infuence of environment
restrict the region of adaptive search for each agent in its ftness landscape.
Its adaptive rules leads to some local peak in the ftness landscape, although
other higher peaks in his ftness landscape can exist as a rule. Restrictions
in behaviour of agents produced via the sum of agent-agent interactions are
working as conditions contributing to (or preventing) the search for optimal
(highest) peak within the ftness landscape. Conditions of realisation
characterises at the macroscopic level of CAS the aggregate effects of agent
interactions on the success of an agents adaptive search.
Adaptive potential is exploited in full if each agent is positioned at the
top of the attractive ftness peak in its ftness landscape. It can be said that
adaptive potential is exploited partially if some agents are not positioned on
these peaks. Effciency of adaptive search within an ensemble of interacting
agents can be described at the macroscopic level as being exploited per an
available adaptive potential ratio.
MSP postulates three universal macroscopic adaptive mechanisms (UMAM)
of CAS work:
1. The accumulation of useful experience (adaptive potential and condi-
tions of realisation) in CAS takes place driven by the adaptive search of
agents within an ensemble. This is the process of learning, training and
fxation of useful experience via mutations. Mathematically this process
can be described as reinforcing feedback process: MSP-potentials
(adaptive potential and conditions of realisation) adaptive.ac-
tivity (adaptive search of agents) increment in MSP-potentials.
(P.1)
2. CAS is capable of maintaining some macroscopic steady state on account
of agent-agent interactions. This property can be described mathemati-
cally as the stabilizing.feedback: temporal equilibrium deviation
from this equilibrium recovery of temporal equilibrium. Macroscopic
steady state of CAS is setted by means of MSP-variables: potential,
conditions and effciency. (P.2)
3. Finally MSP postulates that CAS is capable of controlling distribution
of conditions of realisation between the sub-systems within a CAS.
This property is used by CAS as a means for the maximal realisation
(employment) of accumulated potential. (P.3)
The mechanism (P.3) applied to each CA-sub-system creates multi-level
hierarchical fractal-like structure within comprehensive CAS.
MSP assumes that the previously listed MSP-mechanisms (P.1-P.3) are work-
ing in any real CAS. Attempting to formalise these mechanisms leads to a
new (MSP) platform of CAS-modelling, an intelligent CAS constructed on
MSP postulates (P.1-P.3). MSP is the top-down technique of CAS model-
ling from the integral system to its sub-systems (modules and agents). The
previously listed adaptive mechanisms are working within each sub-system.
In order for each CA-sub-system in comprehensive intelligent CAS operating
on the basis of UMAM-rules (P.1-P.3) necessary impose some restrictions
on agents (modules) behaviour. Laws of behaviour of agents are contained
implicitly in MSP-equations formalizing (P1-P.3) applied to each CA-sub-
system. Evolution of sub-systems (modules and agents) in comprehensive
CAS is the process by means by which whole levels in CAS structure satisfy
requirements of MSP-postulates (P.1-P.3).
The remainder of this chapter is structured in the following way. Section II
contains a brief overview of some results from multi-agent modelling. It is
demonstrated using concrete examples that very dissimilar -models
display one and the same macroscopic properties (*). Section III is devoted to
the discussion of how MAM and MSP platforms can be incorporated. Section
IV introduces the principal positions of MSP-platform within CAS modelling.
Section V contains some results of MSP-modelling and comparisons with
results of MAM-modelling. The Conclusion.contains some ideas of how an
integrated -MSP platform for CAS modelling can be constructed
1
.
Brief.Overview
With the advent of advanced computers, it became possible to track the
dynamics of many MAM-models. The results proved to be unexpected with
consequent outcomes that could not be predicted in advance. CAS does not
contain anything but the agents interacting according to certain rules; how-
Pushno & Bonser
ever, the summary results of such interactions proved to be unpredictable.
The system as a whole showed new, unexpected propertiesthe so-called
emergent properties.
Kauffman and J ohnsen (1991) found that three regimes of macroscopic dy-
namics are possible in their model: chaos, stagnation, and a so-called edge of
chaos. Edge of chaos is the optimal regime of CAS development. Within
this regime, the process of permanent qualitative renewal of the system takes
place. The extinctions of unft species create similar avalanche-like processes.
The power law for frequency distribution of avalanche-sizes indicates that
edge of chaos is a critical state for an ensemble consisting of interacting
agents. This model displays dynamics of punctuated equilibrium character-
ised by long periods of gradual changes followed by short periods of sudden
reconfgurations. However, no mechanism of spontaneous organisation of
the critical state is found in this model.
This drawback was eliminated in the Bak and Sneppen (1993) model of an
ecosystem where the critical state is the attractor and the model shows SOC-
phenomenon. The dynamics in this model has longer periods of relatively
minor and slow qualitative changes, which are interrupted by the bursts of
feverish activity at all levels. Evolution within this model turns out to be a
two-stage process comprising frst of a phase of slow, smooth changes in the
system. This is followed by drastic reconfgurations of qualitative renewalthe
punctuated equilibrium phenomenon.
These two properties of CAS-dynamics (SOC-phenomenon and punctuated
equilibrium) are reproduced in many MAM-models (Amaral & Mayer, 1999;
Sole, Bascompte, & Manrubia, 1996a; Sole & Manrubia, 1996b).
di Collobiano (2002) with colleges analysed the fossil records and concluded
that the collected data indicate the existence of the following macroscopic
properties within the ecosystems of the Earth.
1. Punctuated.equilibrium: The lengthy smooth periods of species de-
velopment are interrupted by short periods of mass-scale extinction of
some species and emergence of new ones (op. cit., p. 31).
2. The power law of frequency distribution of unft species extinctions
against the size of extinctions (Figure 2.6 in op. cit., p. 27) occurs.
3. There exists a delay in recovering biodiversity after mass-scale extinc-
tion of species (op. cit., p. 32).
These observations are evidence of the existence of discontinuous cycles
within the dynamics of real ecosystems, namely: (1) The growth in
diversity of species, followed by (2) a sudden catastrophe and extinc-
tion of a number of species, then (3) a period of delay in origination
of new species, concluding with (4) a rapid recovery and extension of
biodiversity.
4. Finally, analysis of phylogenetic trees shows the presence of hierarchi-
cal structures with fractal properties (Figure 2.2 in op. cit., p. 18).
As far as the ecosystem of the Earth is the real CAS, the properties revealed
during the study of this CAS must have properties that reproduce an intel-
ligent CAS closely resembling reality.
Among the multi-agent models imitating the dynamics of economic and
social CAS, there are some models within which the phases of smooth
changes radically differ in macro-characteristics before and after the proc-
ess of reconfguration. In such models the bursts of activity work similar to
a CAS switching from one work mode to another, these two work modes
alternating.
For instance, Hommess (2002) model of fnancial markets displays the char-
acteristics of the four-phase cycle. The phase of high volatility is replaced
after the burst of reconfguration (switching) by a phase of low volatility.
After a new reconfguration, the process repeats itself. The length of phases
is not constant, but their alternation is strictly regular. Consequently, this
model displays the dynamics of stochastic discontinuous cycles consisting
of two stages, those being the gradual change in the macro-index and two
catastrophe-like switches of the system from one work mode to another.
Epsteins Model of Civil Violence (2002) also shows stochastic discontinuous
cycles in a so-called tension index (Epstein, 2002, Figure 8, p. 7247). The
leaps are caused by avalanche-like transitions of agents from the inactive
state (patience) to an active state (rebellion) and vice versa. As soon as the
tension index exceeds the critical threshold, the system makes a catastrophic
leap to a new equilibrium state.
Kephart (2002) noticed that very dissimilar multi-agent models of market
economies demonstrate the presence of identical dynamic propertiesthe
leaps and cyclic nature of changes in macro-indices. He called these cycles
in market economies price war cycles. He stressed that the reason for such
Pushno & Bonser
dynamics does not lie in the specifcity of particular models, but rather is a
fundamental characteristic of market economies.
we have analyzed and

simulated several different market models. Several
phenomena

appear to be quite generic: not only are they observed in many

different scenarios, but their root causes appear to be very

basic and general
(Kephart,2002;p.7210).
Within the phase of smooth changes in the macroscopic state of the system,
the agents interaction results in the growth in instability of the current equi-
librium. At some point disruption occurs and a new equilibrium is reached;
but it comes at the expense of reconfguration (qualitative renewal) within
the whole system. The reorientation of agents onto new peaks of ftness
(proft) landscapes will be in the form of an avalanche. As a result, a new
confguration emerges with other structures of ftness landscapes and new
agent positions within its landscapes.
Typical price trajectory

consists of linear drops punctuated by sharp discon-
tinuities

up or down. These discontinuities coincide with quick shifts

by all
sellers to a new set of product parameters... Although the intuitive explana-
tions for the sudden price or

price/product discontinuities differ considerably
across the

various models, there is a generic mathematical principle that

explains why this behavior is so ubiquitous and provides insight

into how
broadly it might occur in real agent markets of the

future. Mathematically,
the phenomenon occurs in situations

wheretheunderlyingproftlandscape
contains multiple peaks.

Competition among sellers collectively drives the
price vector

in such a way that each seller is forced down a peak in its

proft
landscape, although its individual motion carries it

up the peak. At some
point,asellerwillfnditbesttoabandon

its current peak and make a dis-
continuous jump to another peak in its landscape. This discontinuous shift in
the price vector

suddenly places all other sellers at a different point in their

ownproftlandscapes.Ifthenextsellertomovefndsitself

near a new peak
in its landscape, it will make an incremental

shift in its price and hence in
the price vector; otherwise,

it will respond with yet another radical shift

(Kephart,2002;p.7211).
Many well-known simple models (sand pile (Bak, Tang, & Wiesenfeld,
1987), percolation (Henley, 1989), forest-fre (Drossel & Schwabl, 1992))
imitating SOC-phenomenon contain some mechanism by means of which the
increase in instability of the current confguration triggers episodic avalanche-
like processes of reconfguration of the system under consideration.
Sudden reconfgurations in MAM-models can be interpreted at the macro-
scopic level as catastrophes of the integral CAS evolution. Sornette (2002)
notes that:
We live on a planet and in society with intermittent dynamics rather than
a state of equilibrium Evolution is probably characterised by phases of
quasistasis interrupted by episodic bursts of activity and destruction (p.
2522).
Interrelation.Between.MAM.and.MSP.Platforms.of.
CAS.Modelling
MAM-MSP interrelation is similar to interrelation between statistical molecu-
lar physics and thermodynamics. Both MAM and MSP describe one and the
same realitycomplex adaptive systems (CAS). MAM attempts to explain
how peculiarities of inner microscopic structure of the system (properties of
agents and behavioural rules) generate the systems macroscopic properties.
MSP attempts to fnd some general regularity in macroscopic properties and
dynamics of ensembles consisting of inter-acting agents. As inter-molecu-
lar interaction and motion of molecules in statistical physics produce some
macroscopic properties of matter, which can be described in thermodynamic
terms just-as-inter-agent interaction and behaviour of agents produce some
regularity in macroscopic properties of the CAS as a whole.
We propose in this essay to apply Jean-Baptiste.Lamarcks.laws of evolu-
tion in order to validate macroscopic regularities of CAS. These regularities
are the properties of ensembles of inter-acting adaptive agents. Each variable
in MSP-platform of CAS-modeling describes some property of an integral
system (an ensemble of interacting agents) irrespective of details of inter-
agent interactions in the system. MSP-potentials (adaptive potential and
Pushno & Bonser
conditions for realisation) play the role of thermodynamic potentials of a
system consisting of many interacting agents.
The MAM-platform is based on the concept of a ftness landscape. Fitness
function describes the ability of agents to survive in perpetually change-
able environments. If the ftness value is larger, the adaptive abilities of the
agent are higher. Therefore, the value of the ftness value depends directly
on adaptive abilities of the agent localised in a defnite position in a ftness
landscape. The adaptive abilities of agents manifest themselves via the rate
of growth in size of an ensemble of agents disposed in a defnite position
in a ftness landscape. Each position in a ftness landscape corresponds to
the defnite genotype of an agent. Mutations in genotype are interpreted as
movement of an agent in a ftness landscape.
The motion in the ftness landscape directed to ftness peak can be interpreted
as an agents adaptive search the best position in a ftness landscape. An
agent uses some criterion in order to choose the fttest position in ftness
landscape. As a rule, an agent tends maximise ftness-value via the mutation
process. Criterion of adaptive search depends on properties of the agent and
its vision. Learning agents are capable of correcting criterion of adaptive
search in accordance with prior experience. If the vision of the agent is
wider, the criterion of adaptive search is more complicated. Agents in real
CAS have as a rule a very wide vision and thus very complex criterion of
adaptive search.
Consider an intelligent CAS in which agents are similar to real (living-like)
agents in a real CAS. It means that we will consider agents that have almost
unlimited vision. Moreover, these agents are capable of correcting criterion
of adaptive search using the whole historical experience of its evolution in
the ftness landscape.
Unlimited vision of a life-like intelligent agent is the agents ability to esti-
mate ftness of each position in its ftness landscape taking into account the
possible losses along a path to this position, and making the best choice.
The change in size of the ensemble of agents moving along a defnite path
depends on ftness values at each point on this path. The possible loss in
population size on the way to a position decreases the appeal of this position.
Only when losses in population on the way to a ftness position exceed some
threshold value does this position become inaccessible to the agent.
Losses in population of ensemble along the way toward ftness peak consist
of two parts: (1) decrement in population on account of small ftness-values
along the path and (2) decrement in population on account of non-optimal
mutations at each point of this path.
The previously stated, a new magnitude which characterises infuence of two
factors may be introduced: (1) ftness value of a defnite position in a ftness
Figure 1. Construction of a complex adaptive system according to MSP-ap-
proach
Notes: Two feedback loops describe the process of adaptation in CAS at the macroscopic
level. The positive feedback (reinforcing feedback process) ensures the accumulation of use-
ful experience (adaptive potential and conditions of its realization) in a System. The
negative feedback (regulating mechanism) stabilizes the temporary equilibrium state of a
System by means of redistribution of potential between realized (exploited) and unrealized
(unexploited) parts. Degeneration of potential and conditions on account of random
destructiveperturbationsofaSystemistheconsequenceofentropyinfuence.
Pushno & Bonser
landscape and (2) the possible losses (costs) in population size on the path
to this position. This will be called magnitude preference.
Preference-value of the defnite position in a ftness landscape depends di-
rectly on ftness value of this position and depends inversely on losses along
the cost-minimizing path leading to this position (Figure 2).
The adaptive search of an agent consists of the following steps.
1. The agent estimates costs (losses) of all-possible paths leading each
point in a ftness landscape.
2. The agent chooses a cost-minimizing path for each point of the ftness
landscape.
3. The agent estimates the preference in each point in the ftness land-
scape.
4. The agent chooses the best future position having the maximal prefer-
ence-value. This position becomes the attractive peak for the agent
under consideration.
5. Agent shifts into a new position in the ftness landscape along a cost-
minimizing path leading to this attractive peak.
The agent, owing to mutations, moves step-by-step along a cost-minimizing
path leading to an attractive peak. The agent recur the procedure of adaptive
search at each step. This mechanism of adaptive search in a ftness landscape
can be specifed by the constriction of the agents vision and defnition in
details of cost- and preference-functions.
As the agent approaches, the attractive peak losses decrease. The position
and the height of this peak are changeable. After the agent culminates this
peak, it keeps in this position in the landscape for some time even though
other higher peaks exist in its ftness landscape. An agent cant reorient onto
another peak because such a choice is burdened by higher losses in population
size on the path towards any new peak. Attainability of the new ftness peak
depends on the losses (cost of attainment) on the path to this peak (Figure
2). Until the adaptive search of the agent takes place in the nearest ftness
peak region the agent evolves gradually.
The point at which preference-value is calculated is the point within geno-
type space. The adaptive potential of this agent is the value of the prefer-
ence-function calculated at the point within the range of possibilities in the
ftness landscape, which represents the highest attainable point (ftness peak)
possible within the range of potential values (Figure 2).
Figure 2. Interrelation between MAM and MSP platforms of CAS-model-
ling
Notes: Fitness-value characterizes the rate of growth of agents population posed in the
defniteplaceofftnesslandscape.Lossesareequaltodecrementinthisrateofgrowthdue
tonon-optimalmutationsonthepathofagentstothedefnitepositioninftnesslandscape.
Preference-valueisftness-valueminuslosses.Thecurrentftness-valueisindexofrealized
potential.Thepreference-valueatthetopofftnesspeakisindexofavailablepotential.
Reorientationofanagentontoanewftnesspeak(levelofcatastrophe)takesplaceif
the preference-value at the top of a new peak is equal or larger than preference-value at
thetopoftheoldftnesspeak.
0 Pushno & Bonser
The quantity of losses along the way toward an attractive peak determines
the conditions of realisation of adaptive potential. The losses along a
path are larger as the conditions of ascent to the top become harder. Losses
along the cost-minimizing path depend directly on the length of this path.
Therefore, an agents adaptive potential and conditions of realisation
increases when the agent approaches the top of its adaptive peak.
The current ftness value of an agent can be interpreted as the agents realised
(employed) adaptive potential. The value of realised potential is lesser than
the value of adaptive potential at each point of a cost-minimizing path.
Each position of an agent in ftness landscape corresponds to the defnite
value of realised adaptive potential (the current ftness value), the adaptive
potential in full (the value of the preference-function at the point of attractive
peak) and the conditions of realisation (the function- formalizing infuence
of the possible losses along the path toward peak) (Figure 2).
Movement of an agent in a ftness landscape consists of two stages: (1) the
gradual approach toward the top of an attractive ftness peak and then (2), a
sharp and sudden reorientation onto the new ftness peak..As the agent ap-
proaches the ftness peak, his gain (increment in preference-value at the point
of attractive ftness peak) tends to zero. From the other hand, the gain from
re-orientation onto a new ftness peak increases because the preference-value
at the point of this new peak grows (Figure 2). Agents perpetually correct
estimations of possible future gains in the process of searching for its next
position in the ftness landscape.
The position of an agent near the current attractive ftness peak is inherently
unstable. Growth of an agents population localised near the defnite ftness
peak is the direct consequence of the high ftness-value in neighbourhood of
this peak. Since the different agents populations are in competitive interac-
tion, the growth of populations will decrease the height of each attractive
ftness peak in each ftness landscape (Figure 3)..As a consequence of this
a new best position in a ftness landscape arises as either a frst or last in-
stance. An agent discovers a new point of maximum of preference-function
in a ftness landscape only when the decrement of the current attractive peak
height becomes large enough. Only when the new attractive peak is found
will an agent abandon the prior peak.
Punctuated equilibrium phenomenon arises as a consequence of the insta-
bility of an agents equilibrium position near a ftness peak. Reorientation
of an agent onto a new ftness peak generates avalanche-like processes of
reorientation in the ensemble of interacting agents. This is stage of CAS re-
confguration. A new structure of ftness landscapes is formed at this point.
This is chaos-like process by which some new CAS structure emerges. At
the macroscopic level, this is reminiscent of a catastrophe in the complex
adaptive systems evolution.
Method.of.Systems.Potential
So, the adaptive potential, , the conditions of realisation, U, and the
realised portion of adaptive potential,
R
, are the basic variables by which
MSP fxes the macroscopic state of CAS.
Figure 3. Dynamics of preference-values in the changeable ftness land-
scape
Notes:Astheheightofftness-peakAdiminishesanewattractiveftnesspeakBarises.
Pushno & Bonser
The adaptation processes lead to the accumulation of adaptive potential and
conditions of realisation in the system. This process takes place through
the adaptive activity at all levels of the CAS-structure, and in all of its sub-
systems. The useful experience (potential and conditions) of CAS is not
concentrated at some particular place within the system, but is dispersed
around all of its constituents.
MSP postulates the presence of three universal macroscopic adaptive mecha-
nisms (UMAM) in any real CAS (P.1-P.3).
Intelligent CAS imitating the effect of these adaptive mechanisms will be
called MSP-systems.
Many researchers (di Collobiano, 2002, p. 31; Lewontin, 1972) insisted that
Darwinism was not suffcient for understanding the process of evolution. The
previously listed properties (*) cannot be explained by Darwins the theory
of natural selection and accumulation of mutation alone.
The MSP-platform of CAS-modelling is based methodologically on the
principles of evolution as offered by J ean-Baptiste Lamarck. He formulated
three macroscopic evolutionary mechanisms: (1) reinforcement of adaptive
abilities in use, (2) extinction of unused adaptive abilities, and (3) inherent
striving of organisms for perfection.
As it often happens in the history of science, these two doctrinesDarwin-
ism and Lamarckismrefectdifferentaspectsofoneevolutionaryprocess.
Struggle for survival and mutations, that is, the concept of Darwinism, are
presented as the basis of the theory of ftness landscapes and multi-agent
modelling. Lamarcks laws of evolution are effective at the level of integral
CAS. Selection and mutations at the level of a multitude of interacting agents
result in CAS developing as an integral whole, just like Lamarck supposed;
the systems used macroscopic adaptive abilities develop, while the unused
ones die off. The interrelation between Lamarckism and Darwinism is the
relation of between the properties of CAS as a united integral system and the
properties of the same CAS as a multitude of interacting agents. Darwinism
describes evolution at the micro-level, the level of interacting agents; La-
marckism describes evolution at the macro-level, that of integral systems.
The adaptive potential of CAS consists of two parts: the portion of potential
realised in adaptive activity,
R
, and the non-realised portion of potential,
D
.
=
R
+
D
. (1)
Lamarcks evolutionary laws regulate the dynamics of these two constituents
of adaptive potential.
Lamarcks.First.Law
Adaptive abilities which are applied in adaptive activities of the system
develop and strengthen.
Exploitedadaptiveabilitiesadaptiveactivityincrementinexploited
adaptive abilities.
The frst of Lamarcks laws can be formalised as reinforcing feedback pro-
cess:
R R
a =
; a >0 (2)
where a is the rate of growth of the realised portion of potential.
Lamarcks.Second.Law
3
.
The adaptive abilities that are not applied in adaptive activities of the system
are degenerate.
Unusedadaptiveabilitiesimpossibilityoftheiruseincurrentactivity
gradual degradation of unused adaptive abilities.
The second of Lamarcks laws can be formalised as follows:
D D
d =
; d >0. (3)
Pushno & Bonser
where d is the rate of degradation of unexploited abilities. This law refects
the action of principle of entropy in CAS.
The.adaptive.activity of an MSP-system is a process of application (exploita-
tion; employment) of adaptive potential. The connection of adaptive activity
with the potential, the conditions of realisation and the realised portion of
potential is determined in MSP with the help of two postulates.
Let A be adaptive activity of a MSP-system in a unit of time.
The.First.Postulate.of.Adaptive.Activity
The adaptive activity of a MSP-system is the source of increment in potential
and conditions of realisation:
c A
+
=
Increment in potential on account of activity.

U q A
+
=
Increment in conditions on account of activity.

c >0, q >0.
Other sources of increase in potential and conditions of realisation in MSP-
system are possible as well. For instance, these values may grow or decrease
on account of external impacts on the system from its interaction with other
systems. Note that these special cases may be organically included in the
general scheme.
The.Second.Postulate.of.Adaptive.Activity.
The exploited potential of MSP-system is larger; therefore, the adaptive
activity of the system is larger.
A =
R
; >0. (4)
As far as the adaptive activity of MSP-system is the employment of the ex-
ploited portion of potential, this postulate is likely to invariably be true for
relevant choices of units of measurement of activity and the potential.
The. current. macroscopic. state,. S,. of the MSP-system is an ordered set
of MSP-variables: (1) potential, , (2) conditions of realisation, U, and (3)
realised portion of potential,
R
:
S (; U;
R
).
a. The frst adaptive mechanism (Lamarcks laws) sets the current equi-
librium state of the MSP-system and the laws governing its change.
Potential, , and conditions of realisation, U, increase with adaptive
activityofaMSP-systemanddecreaseasaconsequenceoftheinfuenceof
the entropy principle.
Mathematically this statement may be represented as follows:
increment on account of decrement on account of
entropy adaptive activity
decr
increment on account of
adaptive activity

c A d
U q A U
=
=
ement on account of
entropy
Stock-fows balance equations

(5)
Multipliers c;d;q; are the rates of the change in the potential and in the
conditions as a result of adaptive activity of the MSP-system and the effect
of entropy.
Differentiating the left and the right parts (1), using Lamarcks laws (2) and
(3) and taking (4) into consideration, it is not diffcult to derive three equa-
tions for the three variables , U and
R
:
( )
decrement on account of
entropy
adaptive activity
adaptive activity

R
R
a d d
U
= +
=
entropy
U
Stock-fows balance equations

(6)
Pushno & Bonser
R R
a =
Lamarcks frst law. (7)

= q , a +d =c . (8)
The second of Lamarcks laws describes the degradation of potential and con-
ditions due to the infuence of entropy. Call parameters a;;;d evolutionary
parameters of MSP-systems. They characterise evolutionary properties of the
system as an integrated whole. Equations (6)-(7) may be solved analytically
if the evolutionary parameters are constant values. Only this simplest case
is considered in this essay.
Figure 1.presents schematically how CAS is constructed at the macroscopic
level within MSP-platform of CAS-modelling. MSP-systems accumulate
their adaptive potential and conditions of realisation (useful experience)
through the channel of adaptive activity and lose them through the channel
of entropy. Adaptive activity is input-fow. The effect of entropy results in
output-fow. The equations representing the balance between the stocks and
fows describe the dynamics of potential and the conditions of realisation.
Equations (6)-(7) do not take into account stochastic perturbations of MSP-
systems.
The above stated, the following defnitions might be introduced.
Density.of.conditions.in MSP-system, z, is the quantity of conditions per
unit of potential.
U
z
. (9)
Effciency of the MSP-system, R, is the rate of exploitation of accumulated
adaptive potential. Effciency equals realised potential per entire potential
ratio:
R
R

. (10)
The following equations respecting density of conditions, z and effciency
R, can be derived from equations (6)-(7):
( ) ( ) 1
R
R
R
a d R
R

= = +

. (11)
( ) ( )
z U
a d R d
z U z
| |
= = + +
|
\ .

. (12)
The following ordinary differential equation follows from equations (11)-
(12):
( ) ( ) ( ) ( ) ( )
1 0
z
R a d z R d z a d R R
( + + + =

. (13)
The solution of this equation determines trajectory of a MSP-system in the
plane (z;R).
Temporary.equilibrium.states.of a MSP-systemcorrespond to points on
the path (13) in the plane (z;R). Each pair (z;R) corresponds to a defnite ray
running through the point of origin in the space (; U;
R
).
According to (11)-(12), the MSP-system having the constant evolutionary
parameters approaches asymptotically to the point in which effciency has
maximal value R =1. Let z
0
be the density of conditions in this point of
maximal effciency. As a consequence of (11)-(12), this value equals:
0
z
a
=
+
. (14)
State (z
0
; R =1) is the global equilibrium state of an MSP-system. Equilib-
rium conditions ( ) ( )
0 0
0 R z z z = =
are fulflled at this point. It is evident that

a MSP-system collapses into its global equilibrium state and remains for an
indefnite time if perturbations of the MSP-system are absent.
What happens to the system if we consider the real situation and take the
existence of random perturbations of the system into consideration? Such
perturbations may be represented by the effect of external perturbation fac-
tors or fuctuations of evolutional parameters.
Under the infuence of such perturbations, the MSP-system deviates from the
equilibrium trajectory (13). The systems dynamics in the plane (z;R) is the
superposition of random wandering on deterministic processes, (11)(12).
As a result, the system approaches toward a point (z
0
;R =1). After that, it
Pushno & Bonser
wanders around this point, moving aside following the perturbations and
returning according to the equations of deterministic process (11)(12). Such
behaviour is very far from the typical picture of real evolutionary process.
The reason for this is that we have not taken into account the systems ability
to maintain (stabilise) its temporal equilibrium state.
b. . The. second. adaptive. mechanism describes the ability of a CAS to
stabilise the temporal equilibrium state. This mechanism is based on the
ability of the system to control the effciency of system performance
through a reduction or extension of the scope of adaptive activity. The
effect of this mechanism may be formalised as a loop of stabilising
negative feedback. Consequently, another equation should be added to
equations (6)-(7).
The.process.of.stabilisation.of.temporal.equilibrium.state.is as follows:
Deviationfromtemporalequilibriumeffectofstabilisingmechanism
return to original temporal equilibrium state.
Stabilizing feedback is represented as regulating mechanism in Figure. 1.
Suppose that the system has gone off its trajectory (13) as the result of a
perturbation. What will be the process of its return into the initial temporal
equilibrium state? Answering this question requires understanding which of
the variables the system is able to control:
U
z =
or
R
R

=
. Stabilisation is
effective only when the process takes place quickly. Neither the potential, ,
nor the conditions of realisation, U, may be changed quickly in response to
random perturbations of the system. These are inertial values. Consequently,
the value
U
z =
is not controlled by the system. The value

R
is the only
variable which is liable to effciently control an MSP-system. The MSP-sys-
tem may react operatively to random perturbations reducing or increasing
the effciency
R
R

=
, thus shifting its state upwards or downwards along

the axis R in the plane (z;R). It is this process which stabilises the temporal
equilibrium state of an MSP-system.
The ability of the system to stabilise its temporal equilibrium is one of the
fundamental properties of any real CAS. For example, stabilisation in living
systems is known as the process of homoeostatic and psychological self-regula-
tion. Stabilisation in the economic system takes place via market mechanisms
of demand-supply-price or investment-savings interdependence.
The choice of mathematical form of the stabilising mechanism in general is
not universal. The equation should be of the following type:
( )
; ;
dR
H R z
dt
= . (15)
Function H(R;z;) depends on values z and as parameters. This function
satisfes the following conditions:
H(R
E
; z) =0; H(R >R
E
; z) <0; H(R <R
E
; z) >0 (16)
( )
; ;
0
E
R z
H
R
<
; R
E
=R(z). (17)
R
E
is the value of effciency in the point of the path (13) having abscissa z.
Assume that parameter does not change. Consider the simplest case of the
equation (15), so-called anti-gradient law:
( )
( )
;
;
R
W R z dR
K K W R z
dt R
= =
; K >>0. (18)
The. stabilizing. function. W(R;z) determines the stability properties of
the MSP-systems temporary equilibrium state. The temporary equilibrium
states (z;R
E
(z)) correspond to the extreme points of function W(R;z) subject
to a fxed z:
( ) ; 0
R E
W R z = . (19)
Points of minimum (maximum and infection) of the stabilizing function,
W(R;z), correspond to stable (unstable) temporal equilibrium states of an
MSP-system. The path of temporal equilibrium states (13), R(z), is a locus
of minimum points of the stabilising function W(R;z) subject to a fxed z.
0 Pushno & Bonser
According to (12) parameter z is changing in with the course of time. If,
having some critical value z, the number of roots, R
E
(z), in the equations
(19), change (for example, two different roots of equation (19) merge into
one root), then the character of stability in the temporal equilibrium state
changes at this value z. The following chapter will show that there are two
critical values of the parameter z in MSP-systems with constant evolution-
ary parameters.
c... The.third.adaptive.mechanism describes the ability of a CAS to maxi-
mise the employed portion of adaptive potential on account for optimal
distribution of conditions between its sub-systems. This mechanism
creates the structure within MSP-system.
The effect of the third adaptive mechanism may be described mathematically
as a conditional extremum problem. Consider an MSP-system consisting of N
sub-systems having potentials
k
and conditions U
k
; k =1;...;N. The aggregate
potentials and conditions of the comprehensive MSP-system is equal to:
1
N
k
k=
=
. (20)
1
N
k
k
U U
=
=
. (21)
When the multipliers are introduced,
k
, which characterise the distribution
of conditions among sub-systems:
U
k
=
k
U,
1
1
N
k
k=
=
. (22)
The conditional extremum problem can be formulated as follows:
R
max,
subject to given ,
{ }
1
N
k
k=
, U.
Consider Lagrange function:
( )
1 2
1 1
; ; ;...; 1
N
N N
k
k k
k k
k
L R U U
= =
| | | |
| |
= +
| |
|
|
\ .
\ . \ .

(23)
where is Lagranges multiplier.
Solution of the equations 0
k
L
=
gives us the following result:

( )
( )

k
k
k
Z k
U
Z Z
dR z
R z Const
dz
= =
= = = (24)
Thus, the distribution of conditions between agents (sub-systems) in a
comprehensive system is optimal if the tangents to the points of path (13) in
which agents (subsystems) are disposed parallel to each other.
In other words, the condition (24) splits the ensemble of sub-systems (agents)
into several groups. The number of groups will be equal to a number of dif-
ferent points of a path (13) in which the tangents to the path are parallel. If
the path has no two differing points with parallel tangents, then all the sub-
systems will have the same state. The ensemble will develop as an integral
whole. If the path has several different points at which the tangents are paral-
lel, then the ensemble will split into several cells creating a structure in the
comprehensive MSP-system. If some cells in their turn are MSP-systems,
then the structure of the second level is possible, and so forth.
In.Summary
Dynamics of the MSP-system is the result of the actions of three universal
macroscopic adaptive mechanisms (UMAM):
a. The frst adaptive mechanism describes the processes of the accu-
mulation of the useful experience (potential and conditions) within the
MSP-system. This mechanism determines the dynamics of the current
state of the system.
Pushno & Bonser

( )
entropy
adaptive activity
adaptive activity

R
R
a d d
U
= +
=
entropy
U
(E1)

R R
a =
(E2)
. Equation.of.trajectory.of.temporary.equilibrium.states.of.a.MSP-
system:
. ( ) ( ) ( ) ( ) ( )
1 0
z
R a d z R d z a d R R
( + + + =

. (E3)
. Equations. of. the. change. of. the. temporary. equilibrium. state. of. a.
MSP-system:
( ) ( ) 1
R
R
R
a d R
R

= = +

. (E4)

( ) ( )
z U
a d R d
z U z
| |
= = + +
|
\ .

(E5)
b. The.second.adaptive.mechanism describes the reaction of the MSP-
system to the effect of external perturbations. This mechanism works
as stabilizing feedback.
. Equation.of.stabilizing.feedback:

( )
( )
;
;
R
W R z dR
K K W R z
dt R
= =
; K >>0. (E6)
c... The.third.adaptive.mechanism.explains the ability of the system to
optimally distribute conditions between the sub-systems:

( )
( )

k
k
k
Z k
U
Z Z
dR z
R z Const
dz
= =
= = = (E7)
This adaptive mechanism is responsible for the generation of the many-level
hierarchical structure of a MSP-system.
Note that the rules of sub-systems interaction cant be stated arbitrarily. Laws
of interaction at the micro-level (between the sub-systems) must have the
condition of coordinated universal mechanisms of adaptation (E1)-(E7) at all
levels of the global system Hence these rules are the conditions necessary for
the coordination of equations (E1)-(E7) made for each of the sub-systems.
Thus, the local properties of the system are interwoven with its global prop-
erties. The laws of the behaviour of sub-systems are implicitly contained in
equations (E1)-(E7).
MSP-Systems.with.Fixed.Evolutionary.Parameters
The equation (E3) can be solved analytically if MSP-system has a constant
evolutionary parameters. The trajectory of the temporal equilibrium states
R(z) (13), consists of two branches (Figure 4):
Upper evolutionary branch: z
()
(R): z =z
0
R C
()
R
(1 R)
1+
;
(25)
Subject to z <z
0
R;
Lower evolutionary branch: z
(+)
(R): z =z
0
R +C
(+)
R
(1 R)
1+
;
(26)
Subject to z >z
0
R;
C
()
>0; C
(+)
>0;
0
z
a
+
;
d
a d
+
; 1
a
< < . (27)

Pushno & Bonser
To specify the constants C
()
and C
(+)
, it is necessary to set two initial positions
of the MSP-system: one at the upper evolutionary branch and the second at
the lower evolutionary branch. The effciency of the MSP-system grows with
time according to the logistic law:
( )
( )
1
1
a d t
R t
b e
+
=
+
;
( )
( )
1 0
0
R
b
R
. (28)
The sign of parameter..determines the shape of function R(z) (Figure 4)..
MSP-systems having positive and negative values have the different dy-
namic and structural properties.
The.First.Type.of.MSP-Systems: > 0
The points of minimum of stabilizing function W(R;z) lie on the upper evo-
lutionary branch and on the lower part of the lower evolutionary branch.
Figure4.Effciencyofadaptation,R(z),asthefunctionofdensityofcondi-
tionsinMSP-systemwithfxedevolutionaryparameters
These are points of stable temporal equilibrium of MSP-system. The points
of maximum for the function W(R;z) lie on the upper part of the lower evolu-
tionary branch (Figure 4). These are points of unstable temporal equilibrium
states of MSP-system.
Supposing, at the initial moment the system was at the point 2 (Figure 5)..It
will move along the lower evolutionary branch until it reaches point 3. While
it is approaching this point, the depth of the potential well W(R;z) in which
the system is situated will decrease (Figure 7). However, this means that the
stability of temporal equilibrium state of the MSP-system is falling as it ap-
proaches point 3. This is the bifurcation point of stabilizing function W(R;z)
at which the minimum point of the function W(R;z) is transformed into a
point of infection (Figure 7). A similar discourse can be made with respect to
point 1 at which the points of maximum and minimum stabilizing of function
W(R;z) converge as well, and the depth of potential well vanishes. At points
Figure 5. Discontinuous evolutionary cycle of MSP-systemwithfxedevo-
lutionary parameters in the plane (z;R)
Pushno & Bonser
1 and 3 the MSP-system, following any small perturbation, will move from
the upper (lower) branch to the lower (upper) branch (Figure 7).
Consequently, there are two critical points at which the temporal equilib-
rium state of the MSP-system becomes unstable. In these points the system
makes a leap from one evolutionary branch to another. Dynamics of an
MSP-systemsubjectto>0.is a sequence of discontinuous cycles (Figure
6, 9, 10, and 12).
Each evolutionary.cycle, 1 2 3 4 (Figure 5) consists of two cata-
strophic jumps and two stages of the gradual motion of the MSP-system along
either lower or upper evolutionary branches. For proper characterisation of
these four phases of the cycle, it is convenient to use the names borrowed
from the business cycle theory: (1) jump downwardscrisis, (2) jump
upwardsrevival, (3) the stage of the gradual development along the lower
evolutionary branchdepression and (4) stage of the gradual development
along the upper evolutionary branchprosperity. Our intuition suggests
Figure 6. Discontinuous evolutional cycle of MSP-system with fxed evo-
lutionary parameters as function of time (perturbations of the System are
very small)
that this is not just an analogy but that in reality the business cycle as such
is based on the evolutionary cycle of the economic CAS.
Instability of the current state of the MSP-system increases in during the
evolutionary cycle. MSP-systems tend toward the global equilibrium (z
0
;
R =1) but this is an unstable equilibrium. This is the unstable attractor of
a MSP-system. Consequently, dynamics of the MSP-system having > 0.
demonstrates self-organised instability phenomenon introduced in the
chapter (Sole et al., 2002).
The leaps occurring in the critical points of the function W(R; z) are the
evidence of mathematical catastrophe. The surface of the catastrophe is
determined by means of the condition:
( )
; ; 0
R
W R z = (29)
Figure 7 shows the surface W(R;z) and the movement of the system around
this surface. At points z
0
and z
1
the MSP-system, under the infuence of the
Figure7.DiscontinuousevolutionarycycleofMSP-systemwithfxedevolution-
ary parameters as motion round the surface of stabilizing function W(R;z)
Pushno & Bonser
stabilising adaptive mechanism (E6), rolls the system into a new the tem-
poral equilibrium state. In these points, the effciency of the MSP-system
changes stepwise on account of a reduction (crisis) or an extension (revival)
of the realised portion of potential.
The simplest approximation of a stabilizing function is four-degree poly-
nomial:
( ) ( ) ( ) ( )
0 1 2
,
R
W R z C R R R R R R =
( ) ( ) ( )
4 3 2
0 1 2 0 1 2 0 1 2 1 2
; ; ; ; ; ;
4 3 2
R R R
W R z A R R R B R R R R R R R C C
| |
= + +
|
\ .
A(R
0
; R
1
; R
2
) =R
0
+R
1
+R
2
; B(R
0
; R
1
; R
2
) R
0
R
1
+R
1
R
2
+R
0
R
2

(30)
C
1
and C
2
>0 are some constants.
Points (z; R
0;1;2
) lie on the evolutionary branches:
z =z
()
(R
2
); z =z
(+)
(R
0;1
) (31)
The evolutionary cycles may be divided into the three large groups: Expan-
sion cycles (potential and conditions, which increase in the long-term
(Figure 14, upper-right and upper-left insets). Contraction cycles (potential
and conditions which decrease in the long-term - Figure 14, lower-right and
lower-left insets). And third, closed cycles (where the changes in potential
and conditions in the long-term are limited).
The results of modelling with the use of equations (E3) - (E6) and (30) are
presented in Figures 9-14. These fgures demonstrate that the length of the
cycle and its different phases are quite sensitive to any small perturbations
to the MSP-system. This follows directly from the fact that the stability of
the systems position on evolutionary branch falls while it is approaching the
points of bifurcation of the stabilising function, and the probability of skip-
ping to the other evolutionary branch grows (Figure 8). The very moments
of transition that determine the length of different phases of the evolutionary
cycle depend on casual factors affecting the MSP-system.
The probability of jumping from the upper evolutionary branches to the lower
branch depends on the average strength of random perturbations (Figure 8).
In the case of small perturbations, the time of the systems life on the upper
branch is protracted. The system mounts nearly to the very crest of the up-
per branch, before transitioning downward. The probability density function
for the events of the fall from the upper branch to the lower one has, as a
density of conditions function, a peak-like form (Figure 15, 16). With the
growth of perturbations, this peak shifts toward lower values of density of
conditions. The probability density function for the events of the systems
shifting from the lower to the upper branch has a form of a peak localised
in the vicinity of the critical point z
1
. With the growth of perturbations this
peak widens and its tail rises up (Figures 15-17).
The probability density of events of skipping from one branch to the other
represents a summary distribution of events of falling downwards and
jumping upwards. In the case of small perturbations within the system, this
resulting distribution has a two-hump form (Figure 15). The frst hump
Figure8.Interrelationofstochasticanddeterministicpatternsindynamics
of MSP-system
0 Pushno & Bonser
Figure 9. Sequence of irregular discontinuous evolutionary cycles of MSP-
Systemwithfxedevolutionaryparameters.Thefollowingvaluesofparameters
wereusedinsimulation:a=0.08;d=0.02;=0.12;v=0.6;z
0
=3;=
1; C
()
=0.05;=0.5
Figure10.DynamicsoftheeffciencyofMSP-Systemwithconstantevolu-
tionaryparameters:a=0.0735;=0.1265;v=0.6;z
0
=3;=1;C
()
=
0.05.StochasticperturbationsofMSP-SystemaremodeledasPDFhaving
=0.5
Figure 11. Dynamics of potential of MSP-System with constant evolution-
aryparameters.ThisisthesamesimulationasinFigure10.
Figure12.SimulationfortheMSP-systemwithweaklyfuctuatingevolution-
ary parameters
Pushno & Bonser
Figure13.DynamicsofpotentialofMSP-systemhavingthefuctuating
evolutionaryparameters(thesamesimulationasinFigure12)
Figure14.SomesimulationsofdynamicsofMSP-systemhavingfuctuating
evolutionary parameters
corresponds to the peak of events created by the fall from the upper branch
down to the lower branch. The second hump matches the peak of leaps
from the lower branch to the upper branch. With the growth of the systems
perturbations, the two-hump distribution is transformed into a distribution
close to the power law (Figures 15-19).
As far as each leap is a catastrophe of qualitative renewal of the system,
which is accompanied by avalanche-like processes of reconfguration, the
power law of frequency distribution for movement from one branch to the
other (Figure 19), shows that some critical state is being established in the
MSP-system. As soon as the average value of external perturbations exceeds
some threshold value, the two-peak-like distribution of fall-events (crises) and
upheavals-events (revivals) (Figure 15) superimposed upon each other (Figure
16) form a power-like law of probability distribution (Figure 17-19).
That is, the MSP-system, with its growth in perturbations, spontaneously
transforms into a critical state, demonstrating the SOC-phenomenon. Each
jump corresponds to some avalanche-like process of system reconfgura-
Figure 15. Non-cumulative frequency distribution of catastrophic jumps as
the function of density of conditions subject to minor perturbations of the
MSP-System; z
1
=2;z
0
=3
Pushno & Bonser
tion. The number of agents engaged within this process depends directly on
the difference between values of effciency before and after the jump. Let
R
max
and R
min
are maximal and minimal values for this difference. If the
value of constants C
(+)
and C
()
tend to zero then R
max
1 and R
min
0.
Consequently, the power law for frequency distribution of avalanche-sizes
may take place within an unlimited range of avalanche-scales. Figure 19 il-
lustrates the linear law within the limited region of avalanche-scales. As the
constants C
(+)
and C
()
are diminish this region is broadened and the system
demonstrates criticality (the power law of frequency of distribution against
avalanche-size within a region of very different scales).
The four-phase discontinuous cycles in which the phases of smooth growth of
the systems effciency alternate with leaping from one evolutionary branch
to the other realise a regime of punctuated equilibrium. The very moments
of quick transition and the smooth development phase length depend on the
effects of random perturbations and are therefore random values.
Figure 16. Non-cumulative frequency distribution of catastrophic jumps as
the function of density of conditions subject to moderate perturbations of
the MSP-System; z
1
=2;z
0
=3
Figure17.Non-cumulativefrequencydistributionofcatastrophicjumpsas
the function of density of conditions subject to large perturbations of the
MSP-system; z
1
=2;z
0
=3
Figure18.Non-cumulativefrequencyofjumps againstthevalueofjump,
R
upper
(z) R
lower
(z)
Pushno & Bonser
Finally, the path of equilibrium states (13) in the MSP-systems with > 0
does not have two different points in which the tangents are parallel. There-
fore, the adaptive mechanism (E7) will concentrate the ensemble of similar
sub-systems (agents) in the neighbourhood of one and the same point of the
evolutionary branch. The ensemble will move as an integral whole, synchro-
nously falling into a crisis phase and in the same way synchronously rising
upwards at the revival phase.
The.Second.Type.of.MSP-System: < 0
Function W(R;z) in this case may be modelled by a polynomial of fourth
order with one point of infection and one point of minimum.
( ) ( ) ( )
2
0 1
,
R
W R z C R R R R =
, C >0 (32)
Figure 19. Non-cumulative frequency of jumps against the value of jump,
R
upper
(z) R
lower
(z), in logarithmic scale
The path of temporal equilibrium (13) has a multitude of pairs of different
points in which the tangents are parallel. The condition of optimality (E7)
may be satisfed here in two different points at once, for instance, at points
1 and 2, as shown in Figure 20. This means that adaptive mechanism (E7)
splits an ensemble of similar MSP-systems with < 0 into two classes: (1)
the class of sub-systems located at the point 1 of the upper branch (lower
class) and (2) the class of sub-systems located at the point 2 of the lower
part of the lower evolutionary branch (upper class) (Figure 20). In other
words, the adaptive mechanism (E7) forms in MSP-systems with < 0 a
cell (class) structure. Points of minimum of the stabilizing function W(R;z)
subject to fxed z coincide with the positions of classes 1 and 2. This is the
simplest variant of class-structure. Evolutionary parameters in real CAS are
changeable values as a rule. Real paths of the temporal equilibrium states are
far more complex curves. As a consequence the adaptive mechanism (E7) is
capable of forming far more complex cell-structures via the decomposition of
Figure20.CatastropheofchaoticrestructuringintheMSP-systemhaving
<0
Pushno & Bonser
the ensemble of interacting agents into many classes (cells). Agents of each
cell are situated in its own potential well of the stabilising function W(R;z).
The stability of the class-structure depends on the deep of potential well of
stabilizing functions in the points where classes are disposed.
The position of cells on the path of the temporary equilibrium states can
change slowly in course of evolution. Therefore, the stability of class-struc-
ture can change as well. Consider the simplest case of MSP-system having
constant evolutionary parameters. Figure 20 illustrates a situation in which
shift of class positions along the equilibrium path leads to a catastrophe of
restructuring. At some critical value z
c

z
0
, the stabilising function is trans-
formed into a level, broad potential well.
If for some reason the effciency of upper class 2 shifts downwards, and
the effciency of lower class 1 shifts upwards, and the condition (E7) is
true for every moment of time, then, having such a shift, the instability of
the class structure will grow, as the depth of potential wells of the function
W(R;z) in which the classes are placed decreases. The agents of one class
will, because of some perturbations, leave their potential well and wander
around the area near the critical value z
c
.
The growing number of wanderings is a sign of initial stage of MSP-sys-
tem reconfguration. The further development in this direction results in the
complete loss of stability in the class structure. The agents of both classes
fall off their potential wells and wander chaotically round the bottom of this
potential well. Collapse of the class structure takes place. This is the catastrophe
of system restructuring. The old structure is destroyed and the system now
searches for a new combination of parts which will form into one structured
whole. This reconfguration takes place according to the scheme: disinte-
grationchaotic wanderingnew assemblage into an integral whole. The
process ends by the formation of a new class structure.
How can these properties of MSP-systems manifest themselves in prac-
tice?
Assume that each agent in a real CAS is the element of two MSP-systems
simultaneously. Being constituents of one and the same integral CAS; all
agents are carriers of the general systemic properties of the comprehensive
big system. As agents having a certain rank within the complex multi-level
hierarchical structure of CAS, agents differ from each other.
The frst type of systemic interrelations is described by way of an MSP-system
with > 0. All agents are disposed in one and the same point on the path (13)
of the MSP-system having > 0. Hence, all agents have one and the same
state within the frame of this MSP-system. This phenomenon is known as
coherence. CAS is capable innately predicting the coherent behavior of all
agents.
The second type of systemic interrelation (structural aspect) is described as
the MSP-system with < 0. Many-scale hierarchical structure exists within
any CAS. Class structure in society and administrative structure of corpora-
tions are formed as the direct result of the struggle of agents for conditions
of survival. Distribution of power and the control over means of subsistence
in society create class structure. This process can be described by means of
MSP-terminology as the optimal distribution of conditions directed to
maximisation of useful output at the level of an integral system.
One and the same.process of qualitative renewal of CAS (reconfguration)
shows itself both as a catastrophic leap within MSP-system having > 0 and
as a catastrophic restructuring of an MSP-system having < 0..Both plans
of systemic interrelations are of great importance for the solution of the
task of effective adaptation of CAS. The mode of best adaptation for CAS
is often called the state of edge of chaos. It is possible to enumerate the
three properties of MSP-systems by which the process of adaptation becomes
most effective.
1.. SOC-phenomenon within of MSP-system having > 0. The transition
into the critical state takes place spontaneously, as soon as the average
value of perturbations exceeds some threshold. The process of qualitative
renewal at the level of the integral system takes the form of an evolu-
tionary cycle. The crisis stage is a manifestation of a reconfguration in
the ensemble of interacting agents.
2. Perpetual renewal of structure within of MSP-system having < 0.
If cells of the structure are posited close to the boundary of the unsta-
ble domain of the stabilizing function, then the composition of cells is
permanently renewed.
3.. There. exists. an. optimal. interrelation. between. the. structural. and.
integral.properties.of.the.system, where the development proves to
be most fexible and effective.
The adaptation of the system is most effective when the structure is renovated
as a result of small crises. The other extreme is a rigid structure, which is
0 Pushno & Bonser
renewed rarely at all levels; this process is accompanied by a catastrophic
global crisis in the system.
These three conditions of the edge of chaos solve one and the same task
of renewing of the system, using the most optimal path.
A certain proportion must exist between introduction of innovations and
preservation of the routines, conditioning the optimal pace of renewal of the
system. CAS in the edge of chaos mode strives to fnd this optimal ratio
between the intensity of innovations and preservation of the routine, when
adaptation proves to be most fexible and effective (Zinchenko, 2006).
Conclusion
Thus, the previously mentioned emergent properties of CAS (*) discovered
within of MAM-platform of CAS-modelling are reproduced in the MSP-plat-
form. Thus there exists a deep, inherent relationship between the and
MSP platforms of CAS-modelling. The combination of these two different
directions within the framework of one approach should make it possible
to advance considerably the understanding of properties and dynamics of
CAS.
Following is a way in which it would be possible to organically synthesise
and MSP platforms.
The MSP-agent.is determined by its position within the class structure in the
MSP-system with < 0. This position may be coded as sequence of symbols
(virtual genome). Development of an MSP-system consists of a sequence
of reconfgurations (evolutionary cycles). Each reconfguration is a change
in the structure and consequently of the positions of MSP-agents within this
structure. MSP-agents pass from one position within the MSP-structure into
another position. This process can be described as the movement of MAM-
agents in a ftness landscape. Consequently, it is possible to organically unite
and MSP platforms into a new more advanced technique of CAS-
modelling. The development of such unifed MAM-MSP platform is a very
diffcult task, but realisation of this task may result in large benefts
5
.
References
Amaral, L. A. N., & Meyer, M. (1999). Environmental changes, coextinc-
tion, and patterns in the fossil record. The American Physical Society,
82(3), 652-655.
Arthur, W. B., Durlanf S. N., & Lane d. A. (Eds.) (1997). The economy as
an evolving complex system II. Proceedings of the Santa Fe Institute:
Vol. XXVII (pp. 1-14). MA: Addison-Wesley, Reading.
Bak P., & Sneppen, K. (1993). Punctuated equilibrium and criticality in a simple
model of evolution. Physical Review Letters, 71(24), 4083-4086.
Bak, P., Tang, C., & Wiesenfeld, K. (1987). Self-organized criticality: An
explanation of 1/f noise. Physical Review Letters, 59(4), 381-384.
di Collobiano, S. A. (2002)..Tangled nature: A model of ecological evolu-
tion. Thesis submitted for the degree of Doctor of Philosophy of the
University of London and the Diploma of Imperial College. Condensed
Matter Theory Group, the Blacket Laboratory, Imperial College of Sci-
ence, Technology and Medicine, London.
Drossel, B., & Schwabl, F. (1992). Self-organized critical forest-fre model.
Physical Review Letters, 69(11), 1629-1632.
Epstein, J . M. (2002). Modeling civil violence: An agent-based computational
approach. Proceedings of the National Academy of Sciences of the U.S.:
Vol.99,Supplement3.SacklerColloquiumonAdaptiveAgents,Intel-
ligence, and Emergent Human Organization: Capturing Complexity
through Agent-Based Modeling (pp. 7243-7250). Washington: National
Academy of Sciences.
Gell-Mann, M. (1994). The Quark and the Jaguar: Adventures in the simple
and complex. New York: W.H. Freeman.
Henley, C. L. (1989). Self-organized percolation: A simpler model. Bulletin
ofAmericanPhysicalSociety,34, 838.
Holland, J . (1995). Hidden order: How adaptation builds complexity. Read-
ing, MA: Addison Wesley.
Hommes, C. H. (2002). Modeling the stylized facts in fnance through simple
nonlinear adaptive systems. Proceedings of the National Academy of
Sciences of the U.S.: Vol. 99, Supplement 3. Sackler Colloquium on
Adaptive Agents, Intelligence, and Emergent Human Organization:
Pushno & Bonser
Capturing Complexity through Agent-Based Modeling (pp. 7221-7228).
Washington: National Academy of Sciences.
Kamke E. (1971). The handbook of ordinary differential equations (4
th
ed.).
Moscow: Nauka.
Kauffman S. A., & J ohnsen, S. (1991). Coevolution to the edge of chaos:
Coupled ftness landscapes, poised states, and coevolutionary avalanches.
Journal of Theoretical Biology, 149, 467-505.
Kephart, J . O. (2002). Software agents and the route to the information
economy. Proceedings of the National Academy of Sciences of the U.S.:
Vol.99,Supplement3.SacklerColloquiumonAdaptiveAgents,Intel-
ligence, and Emergent Human Organization: Capturing Complexity
through Agent-Based Modeling (pp. 7207-7213). Washington: National
Lamarck, J . B. (1809). Philosophie zoologique. Paris.
Lewontin, R. C. (1972). Testing the theory of natural selection. Nature,236,
181-182.
McMillan E. (2004). Complexity, organizations, and change. London and
New York: Routledge.
Pushnoi, G. S. (2003)..Dynamics of a system as a process of realization of
its potential. Proceedingsofthe21
st
International Conference of the
system Dynamics Society, July20-24, New York.
Pushnoi, G. S., (2004a). Method of systems potential and evolutionary
cycles. Working paper presented at inter-disciplinary scientifc forum
(http://www.socintegrum.ru)
Pushnoi, G. S., (2004b, November)..Application of method of systems potential
for analysis of economic system evolution. Paper presented at the Second
Internet Conference on Evolutionary Economics and Econophysics.
Ekaterinburg, Russia: International A. Bogdanov Institute.
Pushnoi, G. S., (2004c, November)..The business cycle model on the basis
of method of systems potential. Paper presented at the Second Internet
Conference on Evolutionary Economics and Econophysics. Ekaterin-
burg, Russia: International A. Bogdanov Institute.
Pushnoi, G. S., (2005, April).. Long-term and short-term dynamics of the
economic system on the basis of systems potential method. Paper pre-
sented at the Third Internet Conference on Evolutionary Economics
and Econophysics. Ekaterinburg, Russia: International A. Bogdanov
Institute.
Sole, R., & Manrubia, S. (1996b)..Extinction and self-organized criticality in
a model of large-scale evolution. PhysicalReview,E54(1), R42-R45.
Sole, R., Bascompte, J ., & Manrubia, S. (1996a). Extinction: Bad genes or
weak chaos? Proceedings of the Royal Society of London B: Biological
Sciences,263(1375), 1407-1413.
Sole, R. V., Alonso, D., & McKane, A. (2002)..Self-organized instability in
complex ecosystems. Philosophical Transactions of the Royal Society
B, 357(1421), 667-681.
Sornette, D. (2002). Predictability of catastrophic events: Material rupture,
earthquakes, turbulence, fnancial crashes, and human birth. Proceedings
of the National Academy of Sciences of the U.S.: Vol. 99, Supplement
1. Sackler Colloquium on Self-Organized Complexity in the Physical,
Biological, and Social Sciences (pp. 2522-2529). Washington: National
Zinchenko, T. N. (2006). Cycles of Russian civilization. Retrieved August 01,
2006, from http://www.socintegrum.ru/forum/viewtopic.php?t=79
Endnotes
1
This paper confnes itself to the most important items of the analysis,
referring Pushnoi (2003; 2004a; 2004b; 2004c; 2005) sources for ad-
ditional details.
2
Sign dot above a variable means time derivative and sign touch
means derivative respect to some variable.
3
This numeration does not coinside with Lamarcks (1809) numeration
of Laws of Evolution from His book Philosophie zoologique.
4
This is J acobis ordinary differential equation. See for example equation
1.250 in Kamke (1971).
5
Author thanks all participants of inter-disciplinary scientifc forum
Socintegrum: http://www.socintegrum.ru for remarks concerning
MSP-approach.
Section II
Important Concepts
Modularty and Complex Adaptve Systems
Chapter.III
Modularity.and.Complex.
Adaptive.Systems
Davd Cornforth, Unversty of NSW, Australa
Davd G. Green, Monash Unversty, Australa
Abstract
Modularity is ubiquitous in complex adaptive systems. Modules are clusters
of components that interact with their environment as a single unit. They
provide the most widespread means of coping with complexity, in both natural
and artifcialsystems.Whenmodulesoccuratseveraldifferentlevels,they
form a hierarchy. The effects of modules and hierarchies can be understood
using network theory, which makes predictions about certain properties of
systems such as the effects of critical phase changes in connectivity. Modular
and hierarchic structures simplify complex systems by reducing long-range
connections, thus constraining groups of components to act as a single com-
ponent. In both plants and animals, the organisation of development includes
modules,suchasbranchesandorgans.Inartifcialsystems,modularityis
Cornforth & Green
usedtosimplifydesign,providefaulttolerance,andsolvediffcultproblems
by decomposition.
Introduction
What is Modularity?
A train consists of an engine and carriages. A tree consists of branches and
leaves. Trains and trees, like innumerable other systems, are built from mod-
ules. Modularity is a structural property of systems that arises when a system
is composed of self-contained groups of elements that behave as a single unit.
Complex systems often contain modules, which increase predictability and
simplify control. Modules can usually be identifed by the pattern of connec-
tions which are stronger and more numerous within modules than between
modules. Modularity has the effect of isolating elements and processes from
one another, and constraining their interactions. Modules isolate functionality
into units that are both reliable and reusable. Complicated problems can be
approached by dividing into smaller problems to reduce the combinatorial
complexity. The formation of modular structures is a crucial mechanism in
the emergence of order in many complex systems. Therefore, modularity is
fundamentally related to the adaptive nature of many complex systems.
The human body provides an example of a natural complex system that
contains a hierarchy of modules. Each cell in the body is a module; the cells
internal component parts and processes are isolated from those of other cells.
The cells themselves are not a homogenous collection. Instead, groups of
cells are specialised and clustered together to form modules. These modules
are recognisable as organs, such as the liver, the heart, and the lungs, as well
as muscles, nerves, and so on. Each organ, or module, can be identifed with
a particular function. The interaction between these modules is well defned.
For example, the heart and kidneys do not interact directly. They perform
specialized functions connected with another entitythe blood. One organ is
responsible for movement of blood around the body. The other is involved in
maintenance of the chemical composition of the blood. The use of modules
in a complex biological system allows cells to operate in an effcient way,
by concentrating on one group of activities.
Modules are the most common mechanism for dealing with complexity,
not only in natural systems, but also in artifcial systems as diverse as or-
ganisational structures and the design of electronic products. A television is
a complex system of electronic components. Such consumer products are
now manufactured in modules using subsystems or circuit boards that can
be easily replaced. Modularity simplifes the tasks of design and repair. The
same advantages are apparent, for example, in traditional J apanese architec-
ture, where room sizes are determined by combinations of rice mats called
tatami, measuring three feet by six feet (a little less than one metre by two
metres). The use of modules allows a reduction in complexity in almost any
context where a system is composed of many interacting parts. Of course,
the organisation of a system into modules involves a cost of implementing
an increased infrastructure. For example, an effort is required to change the
known, working design of a non-modular television to the modular version.
But once in place, the new design has advantages and allows a trade-off in
effort verses effciency.
Modularity plays a key role in complex adaptive systems. In a system that
is changing, the ability to form modules conveys several advantages. First,
modules simplify the process of adaptation by reducing the range of con-
nections. This makes the outcome of any change more predictable and less
prone to errors. In a complex structure, the richness of connections means
that changes often lead to unexpected and sometimes disastrous results.
Complex adaptive systems can avoid such problems by forming modules.
Secondly, modules tend to be stable structures. So forming modules provides
a way of fxing adaptations, thus ensuring that desirable features are retained.
Thirdly, modules provide convenient building blocks, so making it possible
to create large robust systems quickly and easily. We shall see examples of
these effects in later sections.
What is a Hierarchy?
A related concept is hierarchical structure, where a system is composed of
modules that are themselves composed of smaller modules. A hierarchy can
have any number of such levels. The human body has a hierarchy begin-
ning with modules such as the heart, lungs, and kidneys. These modules
are composed of smaller structures; for example, the kidney is composed of
nephrons. These are in turn composed of cells, and cells are composed of
Cornforth & Green
organelles. Like modularity, hierarchical structure seems to be an integral
feature of biological systems. The study of modularity in natural systems
gives many insights that can be applied in artifcial systems. So hierarchies
appear in organisational structures, there are hierarchies in knowledge, such
as the DEWEY catalogue system, and hierarchical structures in manufac-
tured systems. Another example is a power distribution system comprising
power stations, local distribution switchyards, suburban transformers, and
residential dwellings. The unifying principle is effciency and the reduction
of complexity.
Networks.and.Modularity
Networks and graphs are commonly used to represent the structure of com-
plex systems. A network is a system of interconnected entities. An example
is a social network where people are the entities, and they are connected to
each other by relationships such as family relationships or friendships. More
generally, a network is usually described in terms of nodes, which are the
entities, and links, which are the relationships between entities. A graph is a
simple mathematical structure that consists of vertices, or points, to represent
the nodes and uses edges to represent the links that join pairs of nodes. A
network is a graph in which the nodes and/or edges have attributes.
Networks occupy a central position in complexity theory because they un-
derlie both the structure and behaviour of every complex system (Green,
2000). In formal terms, we can obtain a network from the structure of any
system by mapping its elements to network nodes and the relationships (or
interactions) between elements to edges of the network. Likewise, we can
obtain a network from the behaviour of a system by representing all possible
states of the system as network nodes, and representing transitions between
states as edges of the network. These mappings provide a rigorous way of
representing complex structures and behaviour as networks. The implica-
tion is that certain properties of complex systems (e.g., criticality) emerge
as consequences of the underlying network.
Graph theory provides a useful tool for analysing modular and hierarchical
structures in complex systems (Figure 1). The graph in Figure 1(a) represents
a system with seven components. Such a graph could represent a system of
interconnected components in many diverse contexts, for example, cities
connected by roads, or manufacturing workstations connected by the fow
of products and raw materials. In contrast, the graph shown in Fig. 1(b) is
clearly divided into two sections or modules. The number of connections
within each module is relatively high, whereas the number of connections
between the modules is low.
Some basic properties of a graph (Bollobas, 1998) are as follows:
The degree of a vertex is the number of edges meeting at that vertex.
A sparse graph has relatively few edges.
A directed graph has connections that apply in one direction only.
A path is a sequence of consecutive edges. A pair of consecutive edges
shares one vertex.
The length of a path is the number of consecutive edges in the path.
The diameter of a graph is the maximum length that can be found by
considering the shortest path between every pair of vertices. The diam-
eter is fnite only for connected graphs.
A connected graph is one in which every vertex is connected to every
other by at least one path. A connected graph with n nodes has at least
n-1 edges, and its largest possible diameter is also n-1.
A fully connected graph is one in which every vertex is connected directly
to every other. The diameter of a fully connected graph is 1. In a fully
connected graph with n vertices, the number of edges is n(n-1)/2.
(a) (b)
Figure 1. A simple graph (a) and a modular graph (b)
0 Cornforth & Green
A cycle is a closed path, that is, any pair of edges in a cycle shares a
vertex.
The distribution of the degree of vertices in a graph will determine its prop-
erties. This is not constant across the graphs shown in Figure 1, as some
vertices are highly connected, while others have only one connection. Certain
distributions of the degree of vertexes have important properties, as we will
see later.
Random Graphs and Networks
A random graph is a graph in which the edges are distributed at random
among pairs of vertices. This type of graph is least likely to contain mod-
ules, as it has a homogenous structure. The properties of such graphs were
investigated thoroughly by Paul Erds and Alfred Renyi in the 1950s and
60s. In a random graph the probability of any vertex having a given degree
approximates a Poisson distribution, with average degree z =(n-1)p, where p
is the probability of a edge between two vertices and n is the total number of
vertices in the graph. As the value of z increases, a phase transition is reached
where the graph rapidly becomes connected (Erds & Renyi, 1960). The phase
transition also applies in the reverse situation where links are progressively
removed from a network until catastrophic fragmentation occurs. This has
implications for example in the fault tolerance of communications networks.
The average length l of a path between any two vertices is given by:
z
n
l
log
log
=
This means that the average path length can be relatively small even in very
large networks. In communications networks it is desirable to minimise the
length of path that a message must travel between vertices (which are com-
puters), as each edge traversed increases the propagation time for signals,
and increases the probability of errors. Also, the presence of cycles in such
graphs translates to fault tolerance in communication networks: there are
alternative paths for messages to take. Random graphs do not represent the
way natural and artifcial systems are organised. Natural systems tend to
show greater clustering, that is, vertices are grouped together into clusters or
modules. Natural networks are more accurately represented by the models
explained in the next two sections.
Scale Free Networks
A scale free network is a network in which the probability of any vertex hav-
ing a given degree follows a power law (Albert & Barabasi, 2000; Barabasi,
Albert, & J eong, 1999, 2000; Barabasi, Ravasz, & Vicsek, 2001). A few nodes
have many connections, while most nodes have few connections. The vari-
able patterns of connectivity form cliques or modules of highly connected
nodes, with a smaller number of connections between modules. They are
highly tolerant to random attack, but not so tolerant to targeted attack. That
is, if a node is removed at random, the graph is largely unaffected, but if a
node with high connectivity is removed, the graph may become fragmented.
This implies that the Internet, which forms a scale free network, is vulnerable
to attacks on specifc servers, such as malicious attacks by hackers. It also
explains the high degree of robustness exhibited by many complex systems:
individual nodes can be removed without removing the connected property
of the network. However, unlike the catastrophic fragmentations occurring
in random graphs, scale free graphs remain connected even when a large
fraction of edges is removed.
Small World Networks
A small-world network is sparse, highly clustered, and has low diameter
(Watts, 1999; Watts & Strogatz, 1998). These networks consist of a regular
lattice (a tessellation of vertices with uniform connections, for example a
grid) with a number of links added between pairs of vertices chosen at ran-
dom. Such graphs seem to reproduce the small world phenomenon seen in
social networks, where only a short path is required to link any two people.
Such models have implications in many areas, particularly where geography
constrains the interaction between agents, such as in the epidemiology of
disease and the evolution of ecosystems. The probability of vertices having
a given degree follows an exponential distribution.
Cornforth & Green
Trees
A tree is a connected graph with no cycles. Trees are good models of hi-
erarchical systems, for example, the army with its different levels of rank
and rigid route for command and communication. The top level of a tree is
a single node known as the root node. Terminal nodes (those nodes with no
branches descending from them) are called leaves. A tree is called balanced
if each node has the same number of branches descending from it. A subtree
is a subset of nodes that form a tree. Every non-terminal node in a tree forms
the root of at least one subtree. A hierarchy is a directed graph that forms a
tree, and in which all nodes fall on paths leading away from (or to) a single
node, which is called the root of the tree. In a family tree, the descendants
of a single individual form a hierarchy.
Information fow in trees can proceed only from one level to another, as
there are no peer connections. As a tree has no cycles, it affords no fault
tolerance: there is only one path connecting any two nodes, and therefore a
tree is vulnerable to attack. The removal of a single node anywhere results
in the tree being divided into two trees. On the other hand, trees have the
advantage of providing the smallest diameter for a given number of verti-
ces. The maximum number of steps between any two nodes in a regular tree
with L-layers is given by 2(L-1), and L is proportional to the logarithm of
the number of nodes.
Network Modules
Because of their universal nature (see the start of 2.3), networks provide a
convenient way to defne modules mathematically. Within a network, a module
is a connected set of nodes that is richly connected internally, but has only
minimal connections to the rest of the network. In Figure 1(b), for instance,
there are two modules with only a single edge between them.
The previous defnition of a network module is necessarily imprecise. Systems
are not necessarily all modular, or non-modular: they can have degrees of
modularity. This degree can be measured by the modularity coeffcient M.
Various ways of measuring modularity have been proposed (e.g., Muff, Rao,
& Cafisch, 2005; Newman & Girvan 2004). The most obvious approach is
to compare the richness of links within and between modules.
For a known set of modules, the above ratio can be calculated explicitly.
However, a different approach is needed if a measure is needed to identify
emerging modules. One approach is to measure the degree of clustering
(Watts et al., 1998). For any node i in a network of N nodes, the degree of
local clustering is given by the ratio between the number of edges E
i
that exist
between its k
i
neighbours and the maximum possible number of links k
i
(k
i
-
1)/2. The clustering coeffcient C of the entire network is then given by
( )

=
i
i i
i
k k
E
N
C
1
2 1
and the values of C range from 0 to 1. Pimm (1980) proposed a direct measure
of modularity (or compartmentalisation) by looking at modularity about
pairs of nodes. For any pair of nodes i and j, the local modularity m
ij
is the
ratio between the number of neighbours the nodes share in common and their
total number of neighbours, The modularity of the entire network is then the
average of this ratio taken over all pairs of nodes:
=
N
i
N
j
ij
m
N N
M
) 1 (
1
again with values ranging from 0 to 1.
Modularity.in.Mathematics
The idea of building blocks (modules) in mathematics probably began in
prehistoric times with the custom of using fngers for counting. Having ticked
off a set of objects with the fngers of one hand, it is convenient to refer to
that set collectively as (say) a hand or fst. The person doing the counting
could then enumerate larger sets by counting off the number of fsts present.
An abstraction of this idea is the practice of making a mark for each object
and scoring a line through each set of fve marks. The ultimate expression
of this approach to counting is the decimal place system in which each digit
represents entire sets of different size. The number 351, for instance, means
3 sets of 100, plus 5 sets of 10, plus 1 more individual.
Cornforth & Green
Equivalence Relations
An equivalence relation is a pattern that classifes sets of objects as being
related. For example, gender classifes all males as equivalent to one another
and all females as equivalent to one another. An equivalence relation parti-
tions a set of objects into disjoint subsets or classes. That is, different classes
do not overlap. The following mapping of words
cat noun eat verb fat adjective
mat noun sit verb wide adjective
defnes a classifcation of words into their syntactic categories, making nouns
equivalent to one another, and likewise for verbs and adjectives.
Set Theory and Logic
In any system, modules are subsets of the entire system, composed of ele-
ments. These subsets can be defned by logical statements (Figure 2). In
Predicate Calculus, a domain D is a set of objects and an atomic predicate
P(x) is a statement about an element x in D. For instance, if the domain is
the set of people in a village, and P(x) denotes the statement person x is
female, then P defnes a subset of people in the village (i.e. all women)
for which the statement is true. Combining atomic predicates using logi-
cal operators such as AND, OR and NOT makes it possible to defne more
complex statements.
Classifcation
Identifying modules from observational data is the province of data classifca-
tion. Here the problem is to divide a data set into groups of similar things, or
modules. One example is the division of living things into groups including
families, genera, and species. Another example is the division of custom-
ers of a telecommunications company into groups for targeted marketing.
Methods of classifcation can be divided into rule-based and distance-based
approaches.
Rule-based methods identify rules that partition a set of objects according to
particular properties or features. Some examples were shown earlier under
equivalence relations. Perhaps the best-known methods are decision trees.
A decision tree is set of rules that are arranged in the form of a tree graph.
The leaf nodes are outputs (categories in this case). The other nodes contain
rules that take the form of logical tests IF statement A is true THEN follow
branch 1 ELSE follow branch 2. Perhaps the best-known examples are
taxonomic keys for identifying plants and animal species.
Distance-based methods seek to form groups by comparing measures of
distance or similarity between pairs of individuals. A typical measure is Eu-
clidean distance d(x,y) between objects x and y. For instance, if the objects
are (say) survey sites in a landscape, then the attributes might be numbers of
different species found at that site, or environmental parameters.
Algorithms that automatically generate clusters or modules from data are
known as clustering algorithms, and take two possible approaches. Divisive
algorithms carve a set into fner and fner partitions. Agglomerative algo-
Figure2.AVenndiagramshowingvarioussubsetsassociatedwithtwopredi-
cates P(x) and Q(x). The regions represent subsets of elements. The numbers
refertoregionsdefnedbythecompoundpredicates:(1).~(P(x)ORQ(x));
(2).P(x)&Q(x);(3).P(x)ORQ(x)
Cornforth & Green
rithms link together different groups in a series of steps until all items are
joined into one group. At each step, the algorithms link together whichever
pair of groups (or individuals) show the least net separation, according to
the distance measure being used. This procedure leads to a tree with one leaf
node per individual in the set.
Modularity.in.Natural.Systems
Modularity plays a fundamental role in all living systems, from ecosystems
down to the cell. The prevalence of the cell in living things stems from its
effect of concentrating all the ingredients needed to replicate the genome in
one place. Prior to the evolution of the cell, it was probably diffcult to obtain
conditions under which genomic material could replicate at a suffciently
high rate. In multi-cellular organisms, groups of cells interact with one an-
other, and exchange energy, raw products and information. Cells are linked
by well-defned common communication channels: for example, hormones
carried by the blood. They are also linked by the cell membrane dividing
adjacent cells, through which things can pass. Evolution has resulted in highly
complex organisms, and the ability of natural systems to employ modularity
may help to explain how such systems arose. The modular nature of life is
illustrated by the fact that a cell can die and be replaced without affecting
the organism.
Modularity.in.the.Genome
Modularity plays a part in evolution at several different levels of organisa-
tion. At the molecular level, genetic modules promote stability of inherited
characters. Chromosomes themselves are interchangeable sets of characters.
Modular structure is also apparent in proteins. The key structural elements,
known as motifs, are short sequences of amino acids that convey particular
functional properties (Bairoch, 1993).
Modularity is determined by growth processes and appears to be refected in
the organisation of the genes that are involved. In the genome, functionally
related genes tend to cluster together in self-contained modules that code
for particular structures. The best-known examples are the genes that con-
trol gender, which are concentrated on the X and Y chromosomes. During
development, genetic modules are turned on and off by controller genes. For
example, the so-called eyeless gene controls the genes that lead to devel-
opment of the eye. Experiments have shown that if it is made inactive, then
extra eyes can form on the antennae, wings and legs of fruitfies(Halder,
Callerts, & Gehring, 1995). The Hox genes appear to control modules that
determine the basic body plan of individuals (Cohen, 2001).
One of the most fundamental forms of modularity is overall body plan. The
animal designs that exist today all appeared during the Cambrian explosion
(Gould, 1989). In Arthropods, for instance, the basic design is a head fol-
lowed by a number of segments with legs. The number of segments varies
from three in insects to dozens in millipedes.
The.Hierarchy.of.Biological.Taxonomy
Biological taxonomy was created to assist in the identifcation and categorisa-
tion of organisms. There are between two and ten million species of plants,
animals, and micro-organisms either known or assumed to exist. Collectively
these form a complex system, with interactions in the form of evolutionary
descent. In order to simplify the description of species, taxonomy was intro-
duced in the 18
th
century by Swedish botanist Karl von Linn. Taxonomists
divide all living things into categories, or modules, at seven different levels:
kingdom, phylum, class, order, family, genus, and species.
Taxonomic hierarchies are really models of evolutionary change. All species
within a given genus, for instance, are descended from a common ances-
tor. However, missing links and other incomplete information make the
system imperfect. The platypus, for instance, is a fur-bearing, egg-laying,
duck-billed creature, which caused confusion for early taxonomists because
its features suggested that it belonged to several different groups (mammals,
reptiles, birds).
Plant and animal species consist of reproductively isolated groups of organ-
isms and therefore constitute the basic modules of macroevolution. The
genetic modules that control body plans defne the major phyla. They are
now essentially fxed: no new animal phyla have appeared since the early
Cambrian. However, the other levels of taxonomic organisation all mark
major refnements in animal design.
Cornforth & Green
Modularity.in.Plants
Plants make extensive use of modularity. While animals have a fxed number
of parts, plants have open-ended development. Leaf morphology displays
recurring patterns, which are self-similar at different scales (Lindenmayer,
1968). The fractal nature of plant structure is well known. The sub-lobes of
many leaves are quite similar in shape to the larger lobes on the same leaf.
Serrations on sub-lobes are often similar in shape to the large lobes. This self-
similarity is a refection of the fact that plant growth consists of repetition of
modules, like a computer program. This process has been described in models
called Lindenmayer systems (Lindenmayer, 1968). In such models, a simple
grammar defnes transitions between symbols. For example, a symbol for a
stem may be replaced by a symbol for a stem and two symbols for a leaf. Such
models show how simple rules can produce structure with self-similarity, and
they have produced structures strikingly similar to plant forms.
Along the circumference of a growing plant shoot, cells exchange chemical
messages depending on the gradient of concentration of hormones in the cell.
Such interactions may be described by differential equations, as in the model
proposed by the mathematician Alan Turing (1952). The solutions give rise
to standing waves around the shoot. If the hormone is a growth hormone,
then parts of the shoot begin to grow faster than other parts. These parts be-
come the new shoots. The number of shoots that arise from the old shoot is
therefore a number fxed by the parameters of diffusion between cells, and
the mutual effects of the concentrations.
Plants are so reliant on modular structures that insects have evolved to take
advantage of them. Some insects are able to induce the formation of galls by
disturbing the growth sequence of the plants using chemical agents. Galls,
consisting of stunted leaves, serve as a food source or as protection for insect
larvae.
Modularity.in.Embryogenesis
The pattern of early growth, or embryogenesis, is the same for nearly every
vertebrate. The body plan is modular, and these modules can be recognised
very early in the process. For example, the body plan for all organisms is
based on bilateral symmetry. The fertilised cell divides many times. At some
time, a distinction is made between the head and the tail. These are the frst
modules to develop. Then the nervous system develops, and the other inter-
nal organs. Each cell must become a member of a large collection of cells,
forming an organ. In single celled organisms, the cell is very complicated,
as one cell must do everything. In larger organisms, cells can specialise and
reap the benefts in effciency. Although each cell contains the full genetic
code for every function of the organism, it chooses to switch off the majority
of function in order to concentrate on its chosen function.

Modularity.in.Language
Language has a hierarchical structure. Human speech consists of sentences,
words, and various parts of speech. These are some of the modules used to
simplify language. In written language, a book forms a hierarchical struc-
ture, as it is divided into chapters, paragraphs, sentences, words, and letters.
Words are divided into various categories, such as verbs, nouns, and so on.
The indigenous Dyirbal people of Australia have four categories of noun,
one of which has been described as women, fre, and dangerous things
(Lakoff, 1987). It is not necessarily the case that all these things have a
common feature, but that they are related by links made from experiences
or from religious beliefs.
As language is so modular, it is not surprising that the equipment to decode
it is also modular in structure. Neuroscientists have described the brain as
modular in the sense of being composed of cells and layers of cells to divide
the processing. There is evidence to suggest that language is the function of
a specifc part of the human brain (Bates, 1994). Modules deal exclusively
with a particular information type. Such modules include face recognition
in humans and other primates, echo location in bats, or fy detection in the
frog. Such modules may be innate (genetically predetermined) or learned
as a result of interaction with the environment. There is evidence to suggest
that these modules retain plasticity, so that rewiring is possible in the event
of damage.
One way to get an insight into the structure of language, and how it may be
decoded is to attempt to build a machine to understand language. Such ma-
chines are normally constructed by writing a special computer program, and
supplying the language to the computer as digitised stream of letters or as a
digitised audio signal. Many examples have shown that the ability of models
to extract meaning is enhanced by the introduction of modular designs. These
typically focus on dividing speech into phonemes.
0 Cornforth & Green
Modularity.in.Social.Structure
In a social context, modules appear as social groups. The formation of social
groups depends on the richness of interactions within the group. Primates
typically form core groups with mean size of 30 to 40 individuals, which
is maintained through the use of grooming behaviour (Dunbar, 1996). The
development of language in humans allows us to form larger groups with a
mean of 100 to 150. In turn, these larger groups allow greater communica-
tion and cohesion through social gossip. The importance of cooperation as
an evolutionary strategy is well documented.
Patterns of connection and information exchange in society are varied. Some
individuals, such as community leaders, exert a strong infuence upon many
others. Some sections of society make use of hierarchies, for example in
organisational structures. For these reasons, a scale-free network is a good
model of social interactions.
Modularity.in.Artifcial Systems
As modularity is so pervasive in natural systems, it is not surprising that it is
also very common in artifcial systems, since the same issues of complexity
arise, and modularity is able to provide the same answers in different con-
texts. Although these systems are not self-adaptive, they undergo a process
of development as technical knowledge and skills increase over time.
Modularity.in.Software.Engineering
As the complexity of programming tasks has risen, the need has grown for
some way of managing the complex interactions between program instruc-
tions. Modularity has provided that answer for a signifcant part of the history
of programming, and more recently has been refned in the object oriented
paradigm.
The Language Hierarchy
Modern programming development software is designed to make program-
ming easier, by incorporating modularity into the process. A typical integrated
development environment (IDE) has features that can create code for a visual
component such as a button. When the programmer selects a button from a
range of components, the button is displayed on the screen and skeleton code
is automatically generated. That code is compiled into assembler code, which
represents instructions of the microprocessor on which the program will run.
These primitive instructions describe processes in terms of the movement
of data between memory and internal registers of the microprocessor, for
example. Many of these assembler instructions are interpreted by the micro-
processor into the sequence of steps required to perform these operations at
the level of the electronic circuits in the microprocessor. This is termed the
micro code. These different levels of code form a hierarchy, starting from
the instruction by the programmer to create a button, down through several
levels of code to the actions performed by electronic circuits in terms of high
and low levels of voltage.
Structured Programming
The earliest programming languages used statements like goto to direct the
fow of program control. However, as the size and complexity of programs
increased, this type of fow control led to programs that were error-prone
and very diffcult to read, interpret and debug. The introduction of structured
techniques was a great advance, as programs could be divided into modules,
called subprograms, subroutines, procedures, or functions. This allowed
complex programs to be divided up into manageable chunks, so provid-
ing the programmer with the ability to develop programs of arbitrary size.
Solutions to large and complex problems could now be built from modules.
To be successful, this approach requires modules to be small, have a clear
and simple semantics, and be as independent as possible of the context. The
key to managing complexity was the limitation of the number and type of
interactions between modules. The concept of modularity in programming
has since expanded in several ways.
Cornforth & Green
The Object-Oriented Paradigm
Several aspects of modularity are highlighted in the application of object-
oriented languages. This paradigm introduces modularity in two different
contexts, modularity of objects, and modularity of classes.
The whole-part hierarchy extends the concepts of structured programming by
considering a program to be composed of a number of interacting objects. The
object-oriented paradigm increases the control the programmer has over the
encapsulation of instructions, and adds more control over the encapsulation
of data also. This is known as information hiding, where the internal design
of a module is hidden from the outside world. This is, in effect, a way of
limiting the number of connections between modules. In this way, changes
in the internal design of the modules have little effect outside that module.
Information hiding includes the hiding of the internal states, or data, as
well as the hiding of state transition rules, or the internal algorithms. In this
paradigm, the solution to a problem is seen in terms of interacting modules,
called objects. These objects store their internal state in variables, and pos-
sess instructions that are carried out in response to signals from other objects.
The connections between modules are limited by hiding interval variables,
and by strongly defning the type of information that may be passed between
objects in response to signals. Interval variables that are explicitly declared
as private are invisible from outside the object, and cannot be examined by
other objects. Internal instructions are divided into modules called methods,
which are invisible outside the object. Information transfer is limited to data
that can be transferred to the object when a method is invoked, and returned
from the object when a method terminates. The type of data must be explicitly
declared. Figure 3 illustrates these relationships for the structure of a book
using a UML diagram (Larman, 1998). A Book object is a whole object that
is composed of various parts, including Frontmatter, Body, and Annexes. In
turn, the Body of a book is a whole whose parts are the books chapters.
A different kind of object network is the Genspec Hierarchy (general/specifc).
This is based on the concept of inheritance, where the design of a class is
based on another class. A sub class inherits the properties of the super class.
For example, a Book (Figure 4) is based on a higher-level class called Pub-
lication, from which it inherits the properties: Title, Date, and Publisher.
Likewise, a Novel is based upon the Book class, from which it inherits the
properties: Author and ISBN.
Implications of Object-Oriented Programming
The question of how to choose the breakdown of modules involves two costs.
The cost of having too many modules rests on the diffculty of interfacing
all the modules. The cost of having too few modules rest on the diffculty of
developing large modules, since these may be relatively complex. Modern
software development methods include a trade-off between these two costs.
It suggests that a similar trade-off exists in the formation of modular systems
in nature.
The object-oriented paradigm has permeated programming in recent times.
This is surely due to its success as a means of simplifying code, allowing
more effcient code generations and decreasing programming errors. How-
ever, this effciency comes at the price of a tree structure, which is liable to
display the problem of brittleness. This can occur when a phenomenon is
encountered that does not ft into the hierarchy. This problem is often known
as the Platypus Effect, after the animal of the same name that could not easily
be ftted into the existing biological taxonomy.
Figure3.Awhole-partUMLclassdiagramforthestructureofabook
Cornforth & Green
Modularity.and.the.Internet
The Internet is organized as a hierarchy of domains. Computers within an
organisation, for instance, all communicate with a gateway hub, which in
turn exchanges data with a domain controller, and so on. This hierarchy is
refected in the Internet Protocol (IP), which assigns unique numbers to each
computer. So an IP number such as 123.456.789.012 would refer to machine
012 within domain 789. The advantage of a hierarchy is that provides an
effcient compromise between total connectivity and path length. No mes-
sage needs to pass through more than six intermediate computers to reach
its destination. Uniform resource locators (URL), which address item on the
World Wide Web, achieve the same result for Web pages.
The computer network has been shown to have a distribution of connection that
follows a power law: it is a scale-free network (Albert et al., 2000; Barabasi
et al., 2000). Some computers are highly connectedthese are the servers,
gateways and routers. Others have relatively few connections, sometimes only
one. Servers are grouped together within organisations, and form modules that
roughly follow the pattern of organisations. Gateways connect the organisa-
Figure 4. A Genspec class hierarchy relating different types of publica-
tions
tion level network to the outside world, and these gateways in turn form a
larger network that may represent a city. Cities are linked by larger capacity
microwave or cable links. This computer network forms a hierarchy.
Modularity.and.Fault.Tolerance
The interest in living things as analogues for artifcial systems has led to the
consideration of modularity for fault-tolerance. In a living thing, cells can
become defective and die without harming the existence of the organism.
Genes that are crucial to the survival of the organism are duplicated in the
genetic code. Fault tolerance is applied by ensuring an overlapping division
of labour between different modules, as well as different conditions apply-
ing where possible.
In the previous example, the World Wide Web provides information in a
distributed fashion. If any one computer fails, it is very likely that the in-
formation is available somewhere else. A critical feature of this system is
that the servers are geographically distributed, so that they are not subject to
the same conditions. If a power failure occurs in one place, only the servers
located there will fail.
The new discipline of Embryonics (Mange, Sipper, Stauffer, & Tempesti,
2000) promises more fault tolerant machines, which are built from adaptive
modules. Each module has a specifcation like a genetic code, and parts of
this are turned on or off, depending on the required function of the module.
Modules communicate with their peers to perform some collective task, and
are able to sense a faulty neighbour and take over its function. If this system
was in use in robot controllers, for example, any fault could be quickly iso-
lated without human intervention and the corresponding costly shutdown of
an assembly line could be avoided.
Modularity.in.Manufacturing
Modularisation can be found in manufacturing processes of all kinds, as the
ability to make complex systems out of simpler parts reduces costs, allows
easier repair, provides more fexibility of production, and encourages innova-
tion in product development. This is driven by uncertainty in the level of de-
mand, rapidly changing technology, and the increased demand for customized
products. Modularity is closely related to the use of standards. Manufactured
Cornforth & Green
systems are not self-adaptive as some of the preceding examples, as they
rely on a global design process. Nevertheless, they illustrate the application
of modularity in robust design and therefore shed some light on modular-
ity as used by natural systems. Modularity in manufacturing arises in three
contexts: products, processes, and resources.
Product Modularisation
Product modularisation allows products to be assembled from standard,
common parts. This increases the volume of the standard parts, so reducing
costs, and also allows quicker design work. For example, the water pump is
a small, assembled component of a car engine. The same water pump can be
used on many different engines. Not only does this save production costs, as
the volume of water pump production is increased, but it simplifes engine
design. A new engine needs only to conform to the specifed interface for the
water pump. This interface is described in terms of its physical dimensions, its
capacity for pumping water, and its requirement in terms of input rotational
energy. As long as the new engine satisfes these specifcations, the pump will
work with the new engine. In this way, assembled components can be shared
by many different production lines. This reduces the number of different parts
that have to be manufactured and held in stores, thus reducing complexity of
the manufacturing process. Product development time and cost is reduced
by the availability of many modules with well-known properties, which can
be used as building blocks to assemble a new product. The use of modular
components also reduces quality problems, as a module can be thoroughly
tested once, then can be incorporated into many designs.
Process Modularisation
Process modularisation identifes processes that produce standard sub
components, and other processes that add customisations. Processes can be
seen as collections, or modules, of operations performed. For example, the
manufacturing of a motorcar will follow many common processes to build
the basic car, but these may be followed by customisation processes such
as installation of air conditioning or central locking. Process modules can
be re-sequenced so that the processes common to many products are carried
out frst, while the customisation processes are carried out last. The latter
processes may even be delayed until specifc orders are received. By modu-
larizing processes, the manufacturing can be simplifed.
Resource Modularisation
Resource modularisation identifes and exploits modules within resources,
such as staff and workstations. Production teams are commonly composed
of a number of people, often with diverse skills, who collectively form a unit
responsible for certain tasks within the manufacturing process. For example,
in a factory producing motorcars there may be an engine team, or a wiring
team. Resources such as workstations may be formed into modules. For
example, a collection of workstations may form a resource module where
engine assembly is carried out.
Modularity.in.Electronic.Appliances
Modularity has had a huge impact upon electronic engineering, and has fa-
cilitated the widespread availability of cheap consumer products. At the dawn
of electronics, designers who wished to construct an electronic appliance
such as a radio had a number of individual components available to them.
The design of an appliance meant working from frst principles to decide on
a satisfactory arrangement of these components to make the working radio.
This was a major effort involving skilled engineers, and consequently took
time and money. A major innovation in electronics design was the invention
of the integrated circuit, where a commonly used module was fabricated on
a single chip. This meant that a substantial part of the fnished product could
be made using these sub-systems, and consequently a large part of the design
effort could be dispensed with. Engineers were freed from low-level design,
and so attention could be given to the design of more complex systems.
Modularity and Fault-Finding
The traditional approach to fault fnding in electronic circuits is to trace the
signal until it disappears or goes outside its normal behaviour. This is time
consuming and requires a highly trained engineer, and specialist equipment.
The alternative approach is to locate the module containing the fault, and
replace the whole module. This works by a careful design of modules, which
Cornforth & Green
means that each function is associated with a particular module. When the
fault occurs, the module known to be responsible for that function is replaced.
Alternatively, modules can be replaced in turn until the fault is rectifed.
This approach allows much quicker and cheaper repairs, and more economic
manufacturing of the equipment.
Modularity.in.Artifcial Complex Adaptive Systems
The recognition of modular structures within the vertebrate brain have inspired
research into modular neural networks, where part of a problem is solved by
a small network, then other networks are added an their outputs combined
in some way. It is necessary to partition the problem into modules, but this
can be done in some case using clustering algorithms, In other cases, such as
robot controllers, a hierarchical partition is evident in the form of legs, joints
etc. Neural network controllers can be developed in modules as each part of
the mechanism is added to build the robots body. Genetic algorithms can be
used to optimise the organisation and interactions between modules.
An artifcial chemistry is a model composed of molecules and reactions between
them. They have been used extensively to study the origins of life, including
self-organisation. They can be used to model the growth and development
of network structures in evolving complex adaptive systems.
Genetic programming uses a tree based model to evolve computer programs
or complex rules that form solutions to a problem. The crossover operator is
able to copy, move or swap whole sub-trees that form modules of code.
Adaptive.Processes.Leading.to.Modularity
The usefulness of modularity prompts the question of how such modularity
arises, and how it persists. Various adaptive mechanisms contribute to modu-
larity in different contexts. Some of these mechanisms are described next.
Aggregation arises when the autonomous nature of individuals (which tends
to produce disjunctive behaviour) is overcome by connections that serve to
homogenise behaviour. This can lead to individuals coming to the same states,
or updating their states at the same time. Synchrony (see below) is therefore
a possible mechanism for aggregation. To become modular, the aggregation
must produce sub populations within a large population. If all autonomy is
lost, all of the individuals participate in a homogenous collective, rather than
forming modules.
For example, in a forest fre, a tree or a small section of forest can be con-
sidered as an individual. The possible states of an individual are unburnt,
burning, or burnt. Material that is more fammable is likely to change from
unburnt to burning faster than other material. The different areas that ignite
also differ in the length of time they will burn, and therefore in their prob-
ability of igniting neighbouring vegetation. This will lead in turn to pockets
of fre and pockets of unburnt vegetation within the burning forest. These
pockets are modules that are caused by the interaction between areas of dif-
ferent vegetation.
Segregation arises when external constraints are imposed on a complex sys-
tem, splitting it into parts. This type of modularity is caused by conscious
design or physical boundaries. An example of conscious design is the design
of an automobile engine. The interface between engine and gearbox is tightly
controlled. Minor changes to the engine will not require a redesign of the
gearbox, even though these modules are intimately connected. The interface
is fxed, meaning that the connectivity between modules is restricted. Physical
boundaries play a role in the model of speciation proposed by Charles Darwin
(1859). He observed that groups of fnches become separated due to physical
barriers such as separate islands. They consequently experienced different
evolutionary paths, and became distinct species, unable to interbreed. The
connectivity was reduced by the physical separation, then environmental
pressures reinforced the new pattern of connections.
Synchrony refers to the extent to which individuals are able to participate in
shared activities. Such participation can lead to the formation of groups, or
modules. A recent model of this process shows that modules can form due to
clustering of elements that are nearly in synchronisation (Cornforth, Green, &
Newth, 2005). This model uses a modifed one-dimensional cellular automata,
where each cell has its own rate of update, which can be infuenced by the
neighbouring cells (Figure 5a). After some time (Figure 5b), groups of cells
synchronise and produce discrete bands of different patterns, representing
modules of cells. This does not happen when all cells are synchronised, as in
the standard cellular automata model (Figure 5c). In electrical circuits, global
synchronisation by crosstalk is a problem that must be avoided. The abil-
ity to synchronise sub groups of individuals provides a possible mechanism
for emergent modularity, a type of self-organisation often observed in real
00 Cornforth & Green
life systems, such as ant colonies and other social animals. The strength of
temporal connections affects the possibility of synchronisation, so in turn
may determine the maximum size of such systems.
Positive reinforcement is the tendency for an inclination towards some goal
to produce a stronger tendency in the same direction, either through direct
or indirect means. An example of this is Hebbian Learning (Hebb, 1949),
where the synaptic connection between two brain cells that are stimulated
together is strengthened. This makes it more likely that the same two brain
cells will be co-stimulated, which further strengthens the connection. The
phenomenon of positive reinforcement is very common in many natural and
artifcial systems. A further example is Turings model of developing shoots,
Figure 5. Selected images produced by the state activation of a modifed
one-dimensional Cellular Automaton model over time. The cells are organ-
ised as a horizontal row with the two ends connected (to form a loop). Time
goes vertically downwards. The two possible states of each cell are shown
as black and white. Cells are initialised with a random state at the start of
the simulation, at the top of section (a). At a later stage, in section (b), some
cells have synchronised, producing distinct regions. The image in section
(c) represents the same model without the ability to synchronise groups of
cells. This is the standard cellular automata model, where all cells are in
synchrony. These images were produced using on-line software from VLAB
(Green,2006)
(a) (b) (c)
Modularty and Complex Adaptve Systems 0
where the concentration of hormones is eventually maximised in one part of
the growing shoot (Turing, 1952). In a model of social systems, the connec-
tions between individuals holding similar views were strengthened, leading to
the rapid polarisation of individuals into two groups (modules), representing
opposite opinions (Stocker, Cornforth, & Green, 2003).
Co-selection refers to the fact that elements that are close in some sense may
be selected together and therefore after some time these elements may only be
found together. A good example is the phenomena of gene shuffing, which
is believed to play a large part in the formation of clusters of genes. During
reproduction, the genotype is subjected to deletion, copying and insertion
of random segments of DNA. A recent model of gene shuffing showed that
gene clustering is inevitable, and that it leads to modules of genes forming
spontaneously (Green, Newth, Cornforth, & Kirley, 2001).
Holland (1995) proposed a model for the emergence of modules in complex
systems. This model included mechanisms by which agents could interact
with each other via trading, fghting and mating. Agents were described by
a genetic code that represented rules for their interaction, and were placed
in an environment where necessary resources were scarce, so must compete
for these resources. The addition of an adhesion tag to agents allowed them
to join forces to create cartels or modules. The model was able to show the
spontaneous formation of networks between agents, which excluded other
agents and so came to resemble communities or artifcial ecologies.
Conclusion
Modules are ubiquitous in both natural and artifcial systems. Modules are
necessary for the development of structures, where some functionality can
be isolated and protected from change, whilst still allowing other parts of
a system to evolve. Modules allow design to be simplifed, and allow us to
understand the world by breaking down our perception of it into meaningful
chunks. Modularity is fundamentally linked with mechanisms of adaptation in
complex systems. How and why modules form is still the subject of intense
research, but several mechanisms have been identifed, including aggregation
and segregation, and the role of synchrony and positive reinforcement.
0 Cornforth & Green
References
Albert, R., & Barabasi, A. L. (2000). Topology of evolving networks: Local
events and universality. PhysicalReviewLetters,85(24), 5234-5237.
Bairoch, A. (1993). The PROSITE dictionary of sites and patterns in proteins,
its current status. Nucl. Acids. Res. 21, 3097-3103.
Barabasi, A. L., Albert, R., & Jeong, H. (1999). Mean-feld theory for scale-
free random networks. Physica A, 272, 173-187.
Barabasi, A. L., Albert, R., & J eong, H. (2000). Scale-free characteristics
of random networks: The topology of the world-wide Web. Physica A,
281, 69-77.
Barabasi, A. L., Ravasz, E., & Vicsek, T. (2001). Deterministic scale-free
networks. Physica A, 299, 559-564.
Bates, E. (1994). Modularity, domain specifcity, and the development of
language. Discussions in Neuroscience, 10,136-149.
Bollobas, B. (1998). Modern graph theory. New York: Springer Verlag.
Cohen, P. (2001). Monsters in our midst. New Scientist, J uly 2001.
Cornforth, D. J ., Green, D. G., & Newth, D. (2005). Ordered asynchronous
processes in multi-agent systems, Physica D, 204(1-2), 70-82.
Darwin, C. (1859). On the origin of species by means of natural selection.
London: J . Murray.
Dunbar, R. (1996). Grooming, gossip, and the evolution of language. London:
Faber and Faber.
Erds, P., & Rnyi, A. (1960). On the evolution of random graphs. Mathematikai
Kutato Intezet Kozlemenyei, 5, 17-61.
Fodor, J . (1983). Modularity of mind. Cambridge, MA: MIT Press.
Hebb, D. O. (1949). The organization of behavior. New York: J ohn Wiley
Inc.
Holland, J . H. (1995). Hidden order: How adaptation builds complexity.
New York: Addison-Wesley.
Gould, S. J . (1989). Wonderful life: The burgess shale and the nature of
history. W. W. Norton & Company.
Modularty and Complex Adaptve Systems 0
Green, D. G. (2000). Self-organization in complex systems. In T. J. Bossomaier
and D. G. Green (Eds.), Complex systems (pp. 7-41). Cambridge Uni-
versity Press.
Green, D. G., Newth, D., Cornforth, D. J ., & Kirley, M. G. (2001). On
evolutionary processes in natural and artifcial systems. In Whigham,
Richards, McKay, Gen, Tujimura, and Namatame (Eds.), Proceedings of
the 5
th
Australasia-Japan Joint Workshop on Intelligent and Evolutionary
Systems (pp. 1-10), November 19-21, University of Otago, New
Zealand.
Green, D. G. (2006). Artifcial life virtual laboratory (VLAB), Monash
University. Retrieved December 7, 2006, from http://www.complexity.
org.au/vlab
Halder, G., Callerts, P., & Gehring, W. J . (1995). Induction of ectopic eyes
by targeted expression of the eyeless gene in Drosophila. Science, 267,
1788-1792.
Lakoff, G. (1987). Women,fre,anddangerousthings. Chicago, IL: University
of Chicago Press.
Larman, C. (1998). Applying UML and patterns. London: Prentice Hall.
Lindenmayer A. (1968). Mathematical models for cellular interaction in
development I. Filaments with one-sided inputs. Journal of Theoretical
Biology, 18, 280-289.
Mange, D., Sipper, M., Stauffer, A., & Tempesti, G. (2000). Toward robust
integrated circuits: The embryonics approach. Proceedings of the IEEE,
88(4), 516-543.
Muff, S., Rao, F., & Cafisch, A. (2005). Local modularity measure for
network clusterizations. Phys. Rev E, 72, 056107
Newman, M. E. J ., & Girvan, M. (2004). Finding and evaluating community
structure in networks. Phys. Rev E 69, 026113.
Pimm, S. L. (1980). Food Web design and the effect of species deletion.
Oikos, 35, 139-149.
Stocker, R., Cornforth, D., & Green, D. G. (2003). A simulation of the impact
of media on social cohesion. Journal of Advances in Complex Systems,
6(3), 349-359.
Turing, A. (1952). The chemical basis of morphogenesis. Philosophical
Transactions of the Royal Society of London. Series B, Biological
Sciences, 237, 37-72.
0 Cornforth & Green
Watts, D. J ., & Strogatz, S. H. (1998). Collective dynamics of small-world
networks. Nature, 393(6684), 440-442.
Watts, D. J . (1999). Small worlds: The dynamics of networks between order
and randomness. Princeton University Press.
Concept and Defnition of Complexity 105
Chapter.IV
Concept and Defnition of
Complexity
Russell K. Standsh, UNSW, Australa
Abstract
The term complexity is used informally both as a quality and as a quantity.
As a quality, complexity has something to do with our ability to understand
a system or objectwe understand simple systems, but not complex ones.
On another level, complexity is used as a quantity when we talk about some-
thing being more complicated than another. In this chapter, we explore the
formalisation of both meanings of complexity, which happened during the
latter half of the twentieth century.
0 Standsh
Introduction:.Is.Complexity.a.Quality.
or a Quantity?

The term complexity has two distinct usages, which may be categorised simply
as either a quality or a quantity. We often speak of complex systems as being a
particular class of systems that are diffcult to study using traditional analytic
techniques. We have in mind that biological organisms and ecosystems are
complex, yet systems like a pendulum, or a lever are simple. Complexity as
a quality is therefore what makes the systems complex.
However, we may also speak of complexity as a quantitywith statements
like a human being being more complex than a nematode worm, for example.
Under such usage, complex and simple systems form a continuum, charac-
terised by the chosen complexity measure.
Edmonds (1999) performed a comprehensive survey of complexity measures
as part of his PhD thesis, however it has not been updated to include mea-
sures proposed since that time. However, it remains the most comprehensive
resource of complexity measures available to date.
The aim of this chapter is not to provide a catalogue of complexity measures,
but rather to select key measures and show how they interrelate with each
other within an overarching information theoretic framework.
Complexity.as.a.Quantity
We have an intuitive notion of complexity as a quantity; we often speak of
something being more or less complex than something else. However, cap-
turing what we mean by complexity in a formal way has proved far more
diffcult, than other more familiar quantities we use, such as length, area,
and mass.
In these more conventional cases, the quantities in question prove to be de-
composable in a linear way (i.e., a 5 cm length can be broken into 5 equal
parts 1 cm long) and they can also be directly compareda mass can be
compared with a standard mass by comparing the weights of the two objects
on a balance.
However, complexity is not like that. Cutting an object in half does not leave
you with two objects having half the complexity overall. Nor can you easily
compare the complexity of two objects, say an apple and an orange, in the
same way you can compare their masses.
The fact that complexity includes a component due to the interactions between
subsystems rapidly leads to a combinatorial explosion in the computational
diffculty of using complexity measures that take this into account. Therefore,
the earliest attempts at deriving a measure simply added up the complexities
of the subsystems, ignoring the component due to interactions between the
subsystems.
The simplest such measure is the number of parts defnition. A car is more
complex than a bicycle because it contains more parts. However, a pile of
sand contains an enormous number of parts (each grain of sand), yet it is
not so complex since each grain of sand is conceptually the same, and the
order of the grains in the pile is not important. Another defnition used is
the number of distinct parts, which partially circumvents this problem. The
problem with this idea is that a shopping list and a Shakespearian play will
end up having the same complexity, since it is constructed from the same set
of parts (the 26 letters of the alphabetassuming the shopping list includes
items like zucchini, wax, and quince, of course). An even bigger problem
is to defne precisely what one means by part. This is an example of the
context dependence of complexity, which we will explore further later.
Bonner (1988) and McShea (1996) have used these (organism size, number
of cell types) and other proxy complexity measures to analyse complexity
trends in evolution. They argue that all these measures trend in the same way
when fgures are available for the same organism, hence are indicative of an
underlying organism complexity value. This approach is of most value when
analysing trends within a single phylogenetic line, such as the diversifcation
of trilobytes.
Graph.Theoretic.Measures.of.Complexity

Since the pile of sand case indicates complexity is not simply the number
of components making up a system, the relationships between components
clearly contribute to the overall complexity. One can start by caricaturing the
0 Standsh
system as a graphreplacing the components by abstract vertices or nodes
and relationships between nodes by abstract edges or arcs.
Graph theory (Diestel, 2005) was founded by Euler in the 18th century to
solve the famous Knigsberg bridge problem. However, until the 1950s, only
simple graphs that could be analysed in toto were considered. Erds and R-
nyi (1959) introduced the concept of a random graph, which allowed one to
treat large complex graphs statistically. Graphs of various sorts were readily
recognised in nature, from food webs, personal or business contacts, sexual
relations, and the Internet amongst others. However, it soon became apparent
that natural networks often had different statistical properties than general
random graphs. Watts and Strogatz (1998) introduced the small world model,
which has sparked a furry of activity in recent years to measure networks
such as the Internet, networks of collaborations between scientifc authors
and food webs in ecosystems (Albert & Barabsi, 2002).
Graph theory provides a number of measures that can stand in for complexity.
The simplest of these is the connectivity of a graph, namely the number of
edges connecting vertices of the graph. A fully connected graph, however,
is no more complex than one that is completely unconnected. As connectiv-
Figure 1. Various graph theoretic measures for a simple graph. The spanning
trees are shown in the dashed box.
nodes =5
connectivity=6/25
cyclomatic no. =2
spanning trees=4
height (depth)=2
Spanning Trees

"
#
$
%
!

"
"
#
$
%
!
!
#
$
%
"
"
#
$
%
!
!
#
$
%
"
"
#
$
%
!
!
#
$
%
"
"
#
$
%
!
!
#
$
%
ity increases from zero, a percolation threshold is reached where the graph
changes from being mostly discontinuous to mostly continuous. The most
complex systems tend to lie close to the percolation threshold. Another graph
measure used is cyclomatic number of a graph, basically the number of in-
dependent loops it contains. The justifcation for using cyclomatic number
as a measure of complexity is that feedback loops introduce nonlinearities
in the systems behaviour that produce complex behaviour.
Related to the concept of cyclomatic number is the number of spanning trees
of the graph. A spanning tree is a subset of the graph that visits all nodes but
has no loops (i.e., is a tree). A graph made up from several disconnected parts
has no spanning tree. A tree has exactly one spanning tree. The number of
spanning trees increases rapidly with the cyclomatic number.
The height of the fattest spanning tree, or equivalently the maximum number
of hops separating two nodes on the graph (popularised in the movie six de-
grees of separationwhich refers to the maximum number of acquaintances
connecting any two people in the world) is another useful measure related
to complexity. Networks having small degrees of separation (so called small
world networks) tend to support more complex dynamics than networks
having a large degree of separation. The reason is that any local disturbance
is propagated a long way through a small world network before dying out,
giving rise to chaotic dynamics, whereas in the other networks, disturbances
remain local, leading to simpler linear dynamics.
Offdiagonal.Complexity.
Recently, Claussen (2007) introduced a complexity measure called offdiagonal
complexity that is low for regular and randomly connected graphs, but takes
on extremal values for scale-free graphs, such as typically seen in naturally
occurring networks like metabolic and foodweb networks, the internet, the
World Wide Web and citation networks. The apparent ubiquity of the scale-
free property amongst networks we intuitively associate as complex (New-
man, 2003) is the justifcation for using offdiagonal complexity, the other
advantage being its computational practicality.
To compute offdiagonal complexity, start with the adjacency matrix

g
ij
=1 if i and j are connected, g
ij
=0 otherwise (1)
0 Standsh
Let
i
be the node degree of i, and let c
mn
, m n, be the number of edges
between all nodes i and j with node degrees m = (i), n =(j):

, ( ) , ( )
,
mn ij m i n j
i j
c g =

(2)
Then the normalised n-th diagonal sum is

,
,
i i n
i
n
ij
i j
c
a
c
+
=
(3)
The offdiagonal complexity is defned by a Boltzmann-Gibbs entropy-like
formula over the normalised diagonal sums:

ln
n n
offdiag
n
c a a =
(4)
For regular lattices, each node has the same link degree, so c
mn
is diagonal,
a
n
=
n0
and C
offdiag
=0.
For random graphs, most edges will connect nodes with similar link degree
(the characteristic link degree scale), so c
mn
will have a mostly banded struc-
ture, and a
n
0 as n increases. This leads to small non-zero values of the
offdiagonal complexity.
Scale free networks have a power law distribution of link degree, which
leads to the c
mn
matrix having a wide spread of entries. In the case of all a
n

being equal, offdiagonal complexity takes its maximum value as equal to
the number of nodes.
Information.as.Complexity
The single simplest unifying concept that covers all of the preceding consid-
erations is information. The more information required to specify a system,
the more complex it is. A sandpile is simple, because the only information
required is that it is made of sand grains (each considered to be identical, even
if they arent in reality), and the total number of grains in the pile. However,
a typical motorcar requires a whole book of blueprints in its specifcation.
Information theory began in the work of Shannon (1949), who was concerned
with the practical problem of ensuring reliable transmission of messages.
Every possible message has a certain probability of occurring. The less likely
a message is, the more information it imparts to the listener of that message.
The precise relationship is given by a logarithm:

2
log l p =
(5)
where p is the probability of the message, and I is the information it contains
for the listener. The base of the logarithm determines what units information
is measured inbase 2 means the information is expressed in bits. Base 256
could be used to express the result in bytes, and is of course equivalent to
dividing equation (5) by 8.
Shannon, of course, was not so interested in the semantic content of the
message (i.e., its meaning), but rather in the task of information transmis-
sion so instead considered a message composed of symbols x
i
drawn from
an alphabet A. Each symbol had a certain probability p(x
i
) of appearing in a
Figure2.Diagramshowingthesyntacticandsemanticspaces.Twodifferent
messages, having meanings A and B, can each be coded in many equivalent
ways in syntactic space, represented by the sets A and B, The information
or complexity of the messages is related to the size it occupies in syntactic
space by formula (5)

B
Message A Message B
A
L
2
Semantic Layer
B is more complex (or has greater
information than A, because the
set B is smaller than A.
L
1
Syntactic Layer
Standsh
messageconsider how the letter e is far more probable in English text
than the letter q. These probabilities can be easily measured by examining
extant texts. A frst order approximation to equation (5) is given by:

1
1 2 2 1
1
) ( ) log ( ) (
n i i
i
n
p x p x I x x x
=
(6)
This equation can be refned by considering possible pairs of letters, then pos-
sible triplets, in the limit converging on the minimum amount of information
required to be transmitted in order for the message to be reconstructed in its
original form. That this value may be considerably less that just sending the
original message in its entirety is the basis of compression algorithms, such as
those employed by the well-known gzip or PKzip (aka WinZip) programs.
The issue of semantic content discouraged a lot of people from applying this
formalism to complexity measures. The problem is that a message written in
English will mean something to a native English speaker, but be total gib-
berish to someone brought up in the Amazon jungle with no contact with the
English speaking world. The information content of the message depends on
exactly who the listener is! Whilst this context dependence appears to make
the whole enterprise hopeless, it is in fact a feature of all of the measures
discussed so far. When counting the number of parts in a system, one must
make a decision as to what exactly constitutes a part, which is invariably
somewhat subjective, and needs to be decided by consensus or convention
by the parties involved in the discussion. Think of the problems in trying
decide whether a group of animals is one species of two, or which genus they
belong to. The same issue arises with the characterisation of the system by a
network. When is a relationship considered a graph edge, when often every
component is connected to every other part in varying degrees.
However, in many situations, there appears to be an obvious way of parti-
tioning the system, or categorising it. In such a case, where two observers
agree on the same way of interpreting a system, then they can agree on the
complexity that system has. If there is no agreement on how to perform this
categorisation, then complexity is meaningless
To formalise complexity then, assume as given a classifer system that can
categorise descriptions into equivalence classes. Clearly, humans are very
good at thistheyre able to recognise patterns even in almost completely
random data. Rorschach plots are random ink plots that are interpreted by
viewers as a variety of meaningful images. However, a human classifer
system is not the only possibility. Another is the classifcation of programs
executed by a computer by what output they produce. Technically, in these
discussions, researchers use a universal turing machine (UTM), an abstract
model of a computer.
Consider then the set of possible binary strings, which can be fed into a
UTM U as a program. Some of these programs cause U to produce some
output then halt. Others will continue executing forever. In principle, it is
impossible to determine generally if a program will halt or continue on in-
defnitely. This is the so called halting problem. Now consider a program
p that causes the UTM to output a specifc string s and then halt. Since the
UTM halts after a certain number of instructions executed (denoted (p) )
the same result is produced by feeding in any string starting with the same
bits. If the strings have equal chance of being chosen (uniform measure),
then the proportion of strings starting with the same initial (p) bits is 2
- (p)
.
This leads to the universal prior distribution over descriptions s, also known
as the Solomonoff-Levin distribution:

( )
{ : ( ) }
( ) 2
p
pU p s
P s

=
=

(7)
The complexity (or information content) of the description is given by equa-
tion (5), or simply the logarithm of (7). In the case of an arbitrary classifer
system, the complexity is given by the negative logarithm of the equivalence
class size

2 2
( ) lim l og log ( , )
s
C x s N s x
=
(8)
where N is the size of the alphabet used to encode the description and (s,x)
is the number of equivalent descriptions having meaning x of size s or less
(Standish, 2001).
It turns out that the probability P(s) in equation (7) is dominated by the short-
est program (Li & Vitnyi, 1997, Thm 4.3.3), namely

2
( ) log ( ) K s P s C +
(9)
Standsh
(log
2
P(s)<0 naturally) where C is a constant independent of the description
s. K(s) is the length of the shortest program p that causes U to output s, and
is called the Kolmogorov complexity or algorithmic complexity.
An interesting difference between algorithmic complexity, and the general
complexity based on human observers can be seen by considering the case of
random strings. Random, as used in algorithmic information theory, means
that no shorter algorithm can be found to produce a string than simply saying
print ..., where the ... is a literal representation of the string. The algorith-
mic complexity of a random string is high, at least as high as the length of
the string itself. However, a human observer simply sees a random string as
a jumble of letters, much the same as any other random string. In this latter
case, the equivalence class of random strings is very large, close to N
s
, so the
perceived complexity is small. Thus the human classifer defnes an example
of what Gell-Mann calls effective complexity (Gell-Mann, 1994), namely a
complexity that has a high value for descriptions that are partially compress-
ible by complex schema, but low for random or obviously regular systems.
A good introduction to information theoretical concepts for complex systems
studies can be found in Adami (1998).
Information.Theoretic.Graph.Complexity.
In an attempt to bridge the information theoretic approach to complexity with
graph theoretical approaches, Standish recently introduced a coding scheme
for which practical (though still NP-hard) algorithms exist for calculating
the size of the equivalence class of descriptions (Standish, 2005). The in-
tention is to use this method with networks that have a meaning or function
attached, such as metabolic networks, or food webs. A randomly constructed
food web will collapse fairly quickly under ecosystem dynamics to a much
smaller stable food web, so random or regular networks will tend to have a
lower complexity value.
However, when applied to abstract networks, it leads to a perverse result that
regular networks have the high complexities, and the completely connected
network has maximal complexity. This effect can be ameliorated by introduc-
ing a compressed complexity measure, which reduces the complexity measure
of regular networks, and is closer to a Turing complete syntactic language.
Unfortunately, there is no computationally effective algorithm known for
calculating this compressed complexity measure.
Computational.Complexity.and.Logical.Depth
Algorithmic complexity takes no account of the time required to execute
the shortest algorithm. An almanac of tide charts and Newtons equations
of motion plus law of gravity for the Earth-Moon-Sun system contain the
same information, yet the almanac contains the information in a more useful
form for the maritime sailor, as it requires less work to determine when the
tides occur. Computational complexity and logical depth are two concepts
designed to address the issue of a descriptions value.
Computational complexity is defned as the execution time of the algorithm.
Since this is highly dependent on what operations are available to the pro-
cessor, usually only the scaling class of the algorithm is considered as the
input size is increased. Algorithms may scale polynomially, which means the
execution time increases as some power of the problem size (t n
s
), or may
scale faster than this (e.g., exponentially: t n
s
), in which case they have
nonpolynomial complexity. The class of polynomial algorithms is called P,
and the class of nonpolynomial algorithms for which the solution can be
checked in polynomial time is called NP. It is known that all NP algorithms
can be transformed into any other by means of a polynomially preproces-
sor, but it is unknown whether or not P=NP (i.e., whether it is possible to
transform any nonpolynomial algorithm into a polynomial one). This issue
is of great importance, as certain public key encryption schemes depend on
nonpolynomial algorithms to ensure the encryption scheme cannot be cracked
within a practical amount of time.
Bennetts logical depth (Bennett, 1988) is the execution time of the most
highly compressed representation of an object, relative to some reference
machine. It is meant to be a measure of the value of the objectthe almanac
of tide tables has a high value of logical depth compared with the equations
of motion that generate it.
In terms of the observer-based complexity notions introduced in 4, assume
that the observer has a limited amount of computing resources. Perhaps e is
only prepared to spend 5 minutes computing the information, and prior to the
widespread availability of electronic computers this meant that the almanac
was not equivalent to the equations of motion, since it requires more than 5
minutes to compute the tidal information from the equations of motion via
manual paper and pencil techniques. Since the almanac is inequivalent to the
equations of motion in this context, it is clear that the almanac has greater
complexity in this context.
Standsh
Occams.Razor
The practice of preferring a simpler theory over a more complex one when
both ft the observed evidence is called Occams Razor, after William de
Occam:
Entities should not be multiplied unnecessarily.
What is not widely appreciated, is that this strategy is remarkably success-
ful at picking better theories. Often, when tested against further empirical
evidence, the simpler theory prevails over the more complex. A classical
example of this sort of thing is Einsteins General Theory of Relativity. The
key feld equations of general relativity are really quite simple:

8 G T = (10)
Of course unraveling what these equations means for a specifc instance in-
volves a nontrivial amount of 4-dimensional tensor calculus, so the general
relativity computations have high logical depth. Einstein proposed the equa-
tions in this form because they seemed the most beautiful: there were a
large number of alternative formulations that ftted the data at the time. One
by one, these alternative formulations were tested empirically as technology
developed through the 20th century, and found wanting.
However, by what criteria is a particular theory more simple than another.
Goodman (1972) developed a theory of simplicity to put the practice of
Occams Razor on a more rigorous footing. His idea was to formalise the
theories into formal logic predicates, and then count the number of primitive
clauses required to encode the theory.
Solomonoff (1964) developed the concept of algorithmic information com-
plexity in the 1960s as a way of explaining why Occams razor works. He
considered the set of all possible descriptions and computed the probability
distribution that a particular description would be generated by a program
picked at random by the reference machine. His work had some technical
problems that were solved by Levin (1974), which led to the universal prior
distribution (7). Basically, simple descriptions have a much higher probability
than more complex ones, thus Occams razor. The same behaviour is true of
the arbitrary classifer system at equation (8) (Standish, 2004).
While the world is expected to be remarkably simple by the previous argu-
ments, it is also logically very deep (10
10
years of runtime so far!). This appears
to be the result of another poorly understood principle called the Anthropic
Principle. The Anthropic Principle states that the world must be consistent
with our existence as intelligent, reasoning beings (Barrow & Tipler, 1986).
So while Occams razor says we should live in the simplest of universes,
the Anthropic Principle says it shouldnt be too simple, as a certain level of
complexity is required for intelligent life. The simplest means of generating
this level of complexity is by accruing random mutations, and selecting for
functional competence (i.e., Darwinian evolution).
Complexity.as.a.Quality:.Emergence
It is often thought that complex systems are a separate category of systems
to simple systems. So what is it that distinguishes a complex system, such
as a living organism, or an economy, from a simple system, such as a pair
of pliers? This question is related to the notorious question of What is Life?,
however may have a simpler answer, since not all complex systems are liv-
ing, or even associated with living systems.
Consider the concept of emergence (Fromm, 2004; Holland, 1997). We in-
tuitively recognise emergence as patterns arising out of the interactions of
the components in a system, but not implicit in the components themselves.
Examples include the formation of hurricanes from pressure gradients in the
atmosphere, crashes in stock markets, focking behaviour of many types of
animals and of course, life itself.
Let us consider a couple of simple illustrative examples that are well known
and understood. The frst is the ideal gas, a model gas made up of large num-
bers of non-interacting point particles obeying Newtons laws of motion. A
thermodynamic description of the gas is obtained by averaging:
temperature.(T).
is the average kinetic energy of the particles;
Standsh
pressure.(P)
is the average force applied to a unit area of the boundary by the particles
colliding with it;
density.( ).
is the average mass of particles in a unit volume;
The ideal gas law is simply a refection of the underlying laws of motion,
averaged over all the particles:

P

T (11)
The thermodynamic state is characterised by the two parameters T and
.
. The
so-called frstlawofthermodynamics is simply a statement of conservation
of energy and matter, in average form.
An entirely different quantity enters the picture in the form of entropy.
Consider discretising the underlying phase-space into cubes of size h
N
, (N
being the number of particles) and then counting the number of such cubes
having temperature T and density , (T, , N ). The entropy of the system
is given by

( , , ) ln ( , , )
B
S T N k T N = (12)
where k
B
is a conversion constant that expresses entropy in units of J oules
per Kelvin. One can immediately see the connection between complexity
(eq. 8) and entropy. Readers familiar with quantum mechanics will recognise
h as being an analogue of Plancks constant. However, the ideal gas is not a
quantum system, and as h 0, entropy diverges! However, it turns out that
in the thermodynamic limit (N ), the average entropy S/N is independent
of the size of h.
The second law of thermodynamics is a recognition of the fact that the sys-
tem is more likely to move a state occupying a larger region of phase space,
than a smaller region of phase space, namely that (T, , N ) must increase
in time. Correspondingly entropy must also increase (or remain constant)
over time. This is a probabilistic statement that only becomes exact in the
thermodynamic limit. At the syntactic, or specifcation level of description
(i.e., Newtons laws of motion), the system is perfectly reversible (we can
recover the systems initial state by merely reversing the velocities of all the
particles), yet at the semantic (thermodynamic) level, the system is irrevers-
ible (entropy can only increase, never decrease).
The property of irreversibility is an emergent property of the ideal gas, as
it is not entailed by the underlying specifcation. It comes about because
of the additional identifcation of thermodynamic states, namely the set of
all micro-states possessing the same temperature and density. This is extra
information, which in turn entails the second law.
The second example Id like to raise (but not analyse in such great depth) is
the well known Game of Life introduced by Conway (1982). This is a cel-
lular automaton (Wolfram, 1984), in this case, a 2D grid of cells where each
cell can be one of two states. Dynamics on the system is imposed by the rule
that the state of a cell depends on the values of its immediate neighbours at
the previous time step.
Upon running the Game of Life, one immediately recognises a huge bestiary
of emergent objects, such as blocks, blinkers, and gliders. Take gliders for
example. This is a pattern that moves diagonally through the grid. The human
observer recognises this pattern, and can use it to predict the behaviour of
the system with less effort than simulating the full cellular automaton. It is a
model of the system. However, the concept of a glider is not entailed by the
cellular automaton specifcation, which contains only states and transition
rules. It requires the additional identifcation of a pattern by the observer.
This leads to a general formulation of emergence (Standish, 2001). Consider
a system specifed in a language L
1
, which can be called the specifcation, or
syntactic layer (see Figure 2). If one accepts the principle of reduction, all
systems can ultimately be specifed the common language of the theoreti-
cal physics of elementary particles. However, an often believed corollary
of reduction is that this specifcation encodes all there is to know about the
system. The previous two examples shows this corollary to be manifestly
false. Many systems exhibit one or more good models, in another language
L
2
, which can be called the semantic layer. The systems specifcation does
not entail completely the behaviour of the semantic model, since the latter
also depends on specifc identifcations made by the observer. In such a case,
we say that properties of the semantic model is emergent with respect to the
syntactic specifcation.
The concept of good model deserves further discussion. In our previous
two examples, neither the thermodynamic model, nor the glider model can be
0 Standsh
said to perfectly capture the behaviour of the system. For example, the second
law of thermodynamics only holds in the thermodynamic limitentropy
may occasionally decrease in fnite sized systems. A model based on gliders
cannot predict what happens when two gliders collide. However, in both of
these cases, the semantic model is cheap to evaluate, relative to simulating
the full system specifcation. This makes the model good or useful to
the observer. We dont prescribe here exactly how to generate good models
here, but simply note that in all cases of recognised emergence, the observer
has defned a least one semantic and one syntactic model of the system, and
that these models are fundamentally incommensurate. Systems exhibiting
emergence in this precise sense can be called complex.
A school of thought founded by Rosen holds that complex systems cannot
be described by a single best model as reductionists would assume, but
rather has a whole collection of models that in the limit collectively describe
the system (Rosen, 1991). That such systems exist, at least formally, is as-
sured by Gdels incompleteness theorem (Hofstadter, 1979), which shows
that number theory is just such a system that cannot be captured by a fnite
specifcation. He further argues mechanical systems (those that have a fnite
specifcation such as the examples I have previously given) can never be
complex, since the specifcation contains all there is to know about the sys-
tem. However, he implicitly assumes that all models must be perfect (i.e.,
in perfect correspondence with the underlying system), rather than merely
good as I do here. This constitutes a straw man argument, and leads him to
the false conclusion that mechanical systems (e.g., computer simulations)
can never exhibit emergence. The two examples previously presented, which
are perfectly good mechanical systems, are counter-examples to this claim.
Furthermore, the defnition of complex systems presented here is known
to be non-empty, a fact not known of Rosens defnition since no physical
counterpart to Gdels incompleteness theorem is known.
Conclusion
When connoting a quality, complexity of a system refers to the presence
of emergence in the system, or the exhibition of behaviour not specifed in
the systems specifcation. When connoting a quality, complexity refers to
the amount of information needed to specify the system. Both notions are
inherently observer or context dependent, which has lead to a disparate col-
lection of formalisations for the term, and has lead to some despairing of the
concept being adequately formalised. This would be a mistake, as within a
given application, the meaning can be well defned.
An additional diffculty is the combinatorial size of the underlying syntactic
space, which can lead to the intractability of computing complexity. Further-
more, the details of the syntactic layer may be inaccessible, for example the
absence of historical genetic data in the study of evolution from the fossil
record. So being able to establish easy to measure proxies for complexity is
often important, and many proposals for complexity are of this nature.
References
Adami, C. (1998). Introductiontoartifciallife. Berlin: Springer.
Albert, R., & Barabsi, A. L. (2002). Statistical mechanics of complex net-
works. ReviewsofModernPhysics,74, 47.
Barrow, J . D., & Tipler, F. J . (1986). The anthropic cosmological principle.
Oxford: Clarendon.
Bennett, C. H. (1988). Logical depth and physical complexity. In R. Harkin
(Ed.), The universal turing machine: A half century survey (Vol. 1, pp.
227-258). Oxford: Oxford UP.
Bonner, J . T. (1988). The evolution of complexity. Princeton: Princeton UP.
Claussen, J . C. (2007). Offdiagonal complexity: A computationally quick
complexity measure for graphs and networks. PhysicaA,375,365-373.
arXiv:q-bio.MN/0410024.
Conway, J . H. (1982). What is life? In E. Berlekamp, J . H. Conway, & R.
Guy (Eds.), Winning ways for your mathematical plays (Vol. 2, Chapter
25). New York: Academic.
Diestel, R. (2005). Graph theory (3rd ed.). Berlin: Springer.
Edmonds, B. (1999). Syntactic measures of complexity. PhD thesis, Uni-
versity of Manchester. http://bruce.edmonds.name/thesis/, Complexity
Bibliography: http://bruce.edmonds.name/compbib/.
Standsh
Erds, P., & Rnyi, A. (1959). On random graphs. Publ. Math. Dubrecen,
6, 290-291.
Fromm, J . (2004). The emergence of complexity. Kassel: Kassel UP.
Gell-Mann, M. (1994). The Quark and the Jaguar: Adventures in the simple
and the complex. London: Little Brown Co.
Goodman, N. (1972). Problems and projects. Indianapolis, IN: Bobbs-Mer-
rill.
Hofstadter, D. R. (1979). Gdel, Escher, Bach: An eternal golden braid. New
York: Basic Books.
Holland, J . (1997). Emergence: From chaos to order. Reading, MA: Ad-
dison-Wesley.
Levin, L. A. (1974). Laws of information conservation (non-growth) and
aspects of the foundation of probability theory. Problems Inform. Trans-
mission,10, 206-210.
Li, M., & Vitnyi, P. (1997). An introduction to Kolmogorov complexity and
its applications (2nd ed.). New York: Springer.
McShea, D. W. (1996). Metazoan complexity and evolution: Is there a trend?
Evolution,50, 477-492.
Newman, M. E. J . (2003). The structure and function of complex networks.
SIAMReview,45(2), 167-256. arXiv:cond-mat/0303516.
Rosen, R. (1991). Life itself. New York: Columbia UP.
Shannon, C. E. (1949). The mathematical theory of communication. Urbana-
Champaign, IL: University of Illinois Press.
Solomonoff, R. J . (1964). A formal theory of inductive inference, part 1 and
2. Inform. Contr., pages 1-22,224-254.
Standish, R. K. (2001). On complexity and emergence. Complexity Interna-
tional, 9. arXiv:nlin.AO/0101006.
Standish, R. K. (2004). Why Occams razor? Foundations of Physics Letters,
17,255-266. arXiv:physics/0001020.
Standish, R. K. (2005). Complexity of networks. In Abbass et al. (Eds.), Recent
advances in artifcial life, volume 3 of Advances in Natural Computation,
pages 253-263, Singapore. World Scientifc. arXiv:cs.IT/0508075.
networks. Nature,393(6684), 409-10.
Wolfram, S. (1984). Cellular automata as models of complexity. Nature,
311,419-424.
Key.Terms.
Algorithmic.Complexity: Length of shortest program capable of generating
the description (also known as Kolmogorov complexity).
Anthropic.Principle: The statement that the properties of the universe must
be such as to permit our existence as human observers.
Cellular.Automaton: A grid of cells, each of which can be in a fnite number
of states. The states of each cell depend only on the states of a neighbour-
hood of cells at the previous timestep, known as a transition rule.
Church-Turing.Thesis:.The proposition that all computable functions can
be computed on a Turing machine.
Classifer System: A system that classifes a set of objects into a discrete set of
classes. Formally equivalent to a map into the set of whole numbers.
Complexity.(Quality):.The quality of possessing emergent properties.
Complexity. (Quantity): The amount of information a particular system
represents to the observer.
Computational. Complexity:. The computational cost of executing an al-
gorithm.
Emergence:.The phenomenon of emergent properties.
Emergent.Properties:.Properties of a system at the semantic level that are
not entailed at the syntactic level.
Entail:.To logically imply something.
Entropy: Logarithm of number of syntactic states equivalent to a given
semantic state. It is related to information (I) by S+I=S
max
where S
max
is
the maximum possible entropy of the system.
Equivalence.Class: A set of objects that are equivalent under some mapping,
i.e.,
{ : ( ) ( ')} ' x e x e x x =
.
Game.of.Life: A well known cellular automaton, with two states per cell,
and a particular transition rule.
Standsh
Graph.Theory: Mathematical theory of objects consisting of atomic nodes
linked by connections called edges.
Gdel.Incompleteness.Theorem: No fnite set of axioms can prove all pos-
sible true theorems of number theory.
Information:. The amount of meaning in any description; formally given
as the logarithm of the proportion of syntactic space occupied by the
description.
Link.Degree:.The number of edges a node has in a graph.
Logical.Depth:.Execution time of the most compressed representation of
an object.
Newtons.Laws.of.Motion:.Laws of ideal point particles: the acceleration
a particle experiences is proportional to the force acting on it, which
is a function of the positions and velocities of the particle and the en-
vironment.
Occams.Razor: A statement of the practice of preferring a simpler theory
over a more complex one.
Scale. Free: A scale free distribution has infnite mean. A power law dis-
tribution is a common scale free distribution. A scale free process is a
stochastic process obeying a scale free distribution.
Second.Law.of.Thermodynamics: Entropy can only increase, or remain
constant in a closed system; it can never decrease.
Semantic.Level.(or.Space):.The space of meanings for any description.
Syntactic.Level.(or.Space): The language in which a description is speci-
fed: letters of the alphabet, genetic code, laws of theoretical physics,
as appropriate.
Turing.Machine: A formal model of a computation.
Universal.Turing.Machine: A formal model of a computer. Is capable is
simulating any Turing Machine with appropriate input.
Section III
Computing Perspectives
Thrsson
Chapter.V
Emergence.of.Creativity:
A.Simulation.Approach
Hrafn Thorr Thrsson, Reykjavk Unversty, Iceland
Abstract
This chapter presents a theory of natural creativity and its relation to cer-
tain features of intelligence and cognitive faculties. To test the theory we
employ simulated worlds of varying complexity, inhabited by creatures with
a genetically-evolving mental model. Plan-making strategies are compared
between creatures in each of these worlds. The results show that creative
behaviors are governed by the worlds structural coherence and complex-
ity.Inlightoftheresultswepresentanewdefnitionofcreativity,proposea
theory for why creativity evolves in nature, and discuss creativitys relation
to perception, goals, logic, understanding, and imagination. Creativity has
beenadiffcultconcepttodefneanditsexactrelationshipwithintelligence
remains to be explained. The theoretical framework presented is proposed
Emergence of Creatvty
as a foundation and tool for furthering understanding of natural creativity
andtohelpdevelopcreativeartifciallyintelligentsystems.
Introduction
Creativity is an important capability of humans. It gives us power beyond
the reach of any other intelligent system known and separates us from the
rest of the animal kingdom. Through creativity, the human race has learned
to build complex technology far beyond that of any other species on this
planet (Figure 1). Although creativity is an enormously important aspect
of our existence, few artifcial intelligence (A.I.) studies have been directly
aimed at understanding the general underlying structure, or nature, of creativ-
Figure 1. Architectural constructs are perhaps not considered prime examples
of creative output, yet are based on numerous creative insights without which
they would not exist. Compared to some of the structures considered very
creative in the animal kingdom, such as bird nests or ant-hills, the difference
of creative capacity is obvious.
Thrsson
ity. The emphasis in A.I. studies has so far mainly been on building logical
systems and it seems that creativity has for the most part been considered a
side-product of logic.
Often associated with genious, creativity has thus been a highly visible part
of human civilizationcreative individuals set new standards and goals for
humanity by pioneering physical constructions and scientifc concepts. Al-
though prime examples of creativity are generally extraordinary, it is quite
obvious that creativity is also at work when it comes to the more mundane,
such as plumbing, planning the route to your workplace, or fnding a way to
pour coffee when one hand is holding a telephone and the other a cat. Any
plan to manipulate our environment or fulfll our goals is created in accor-
dance to specifc circumstances, whether the plan is to portray a woman on
canvas or avoid spilling your coffee on your cat.
Creativity affects many features of intelligence (particularly in humans),
making it very diffcult to formally defne as a concept; one cannot say that
creativity is defned only by the novelty of the phenomena it produces, for it
must be logical to some extent as well; randomness is generally not considered
creative. It is also subtly erroneous to state that something is creative if it is
novel and logicalbecause the logicality of a novel act is closely tied with
that of the goal it belongs to, and thus an act can be extremely creative and
logical within the confnes of its goals illogicality. To complicate things even
more, intentionality of a creative act is a determining factor of its validitywe
wouldnt normally perceive a person as creative if a certain action of theirs
had led to a coincidental discovery of a new scientifc concept (although such
events may of course occasionally happen).
In light of the vast spectrum of human creative capabilities, noteworthy at-
tempts at computational replication are often governed by an anthropocentric
viewpoint, pushing research away from associating creativity with that of
less capable animals that might provide us with a better understanding of
the innate nature of its underlying mechanisms. An overemphasis on human
capabilities may well be partially to blame for the fact that a wholesome
defnition and explanation of creativity has not yet reached a general con-
sensus. In this chapter, we introduce a new simulation model consisting of
virtual creatures within emergent environments and propose new methods
with which we can begin exploring how creativity can be defned and identi-
fed in simpler animals.
The chapter is organized as follows: We begin by reviewing the background
of creativity studies in various scientifc disciplines. We then state our hy-
potheses in more detail and describe a simulation framework created for
exploring how creativity exhibits itself in agents evolving within emergent
environments. Then we describe an experiment based on the platform, and
discuss the results. After which we discuss the theory in detail, in light of the
experimental results and concepts of emergence, providing a new defnition
of creativity. We end the chapter with future research possibilities.
Background
Although some progress has clearly been made in research on creativity over
the last 20 years, the concept has remained extremely hard to defne and un-
derstand. By 1988, over 50 different defnitions of creativity had already been
proposed (Taylor, 1988). A common thread among many of them is a primary
distinction between creativity displayed in an artifact (evaluated by society),
and processes that might result in a creative product (Gero, 1996), or artifact
(e.g., poems, paintings, architecture, music). The characteristics of creative
artifacts are generally considered to be (1) novelty, (2) unconventionality or
unexpectedness, and (3) value of the product (value to whom remains un-
defned). These have often been the pillars from which creative artifacts are
evaluated (Boden, 2004; Gero, 1996; Liu, 1996; Wiggins, 2006). However,
there exist no empirical evaluation methods for creativity or creative prod-
ucts, exploring processes by identifcation of the resulting products remains
diffcult. Another common view of human creativity is seeing it as merely a
recombination of old ideas, which deserves special attention here. The idea
has received signifcant opposition, especially in that it does not capture the
entire spectrum of human creativity. Boden (1995) has argued that:
Creativenoveltiesareofsignifcantlydifferenttypesthemostinteresting
of which lie beyond combination-theory. Many creative ideas are surprising
not because they involve some unusual mix of familiar ideas, but in a deeper
way. They concern novel ideas which not only did not happen before, but
whichin a sense that must be made clear (and which combination-theory
cannotexpress)couldnothavehappenedbefore(p.124).
0 Thrsson
Moreover, combination-theorists do not explain how novel combinations come
about in the frst place. While recombinations can contribute to creativity,
there is more to the story.
Creativity affects behavior to such a high degree that its been hard to discern
the differences and correlations of creativity and intelligence. Runcos (2004)
review of creativity research covering the past 20 years mentions many pro-
cesses studied in cognitive sciences that relate to creativity:
Basic cognitive processes that have been studied include memory (Pollert,
Feldhusen, Van Mondfrans, & Treffnger, 1969), attention (Martindale &
Greenough,1973),andknowledge(Mumford,Mobley,Uhlman,Palmon,&
Doares, 1991, Rubenson & Runco 1995). Cognitive research has also focused
ontactics,strategies,metacognition(Adams,1980,Root-Bernstein,1988,
Runco,1999c),andintellectualskills...(p.667).
Considering the high number of cognitive features that creativity has been
associated with, it has been suggestedand indeed seems likelythat
creativity is not a specifc mechanism of the brain but rather an aspect of
human intelligence in general (Boden, 2004)..Humans are extremely com-
plicated machines, and admittedly more so than other animals. However,
if creativity is tightly coupled with intelligence it must be considered that
creativity is also evident in the intelligence of simpler animals and that the
tendency to associate creativity with (extraordinary) human abilities may
have been overemphasized. Crows have exhibited an ability to bend strings
of wire into various shapes to use as tools for retrieving food from places
that are diffcult to get to (Weir & Kacelnik, 2006). Research indicates that
this ability is not coincidental, and that the crows did not solve problems by
repeating previously learned actions, but by generating new plans to solve
each specifc problem. We can assume, if creativity is a phenomenon evident
in all intelligences, that its evolutionary roots are tightly related to planning
mechanisms (e.g., as used by the crow to retrieve food). All non-reactive
animals have some mechanisms for dynamic planning, and as such it could
be that creativity shares common evolutionary roots in all animals (well
come back to this issue later).
Some attempts have been made to explain the origins of creativity in terms
of evolution (most of which are presented in relation to human creativity),
and it has been suggested that exploring creativitys evolutionary roots might
give us a holistic picture of how and why it emerges in relation to other
processes (Gabora, 2005; Thrisson, 2004). Carruthers (2002) has proposed
that the function of the extensive creative play of children, also evident in
the behavior of other mammalians where the young engage in pretend-play
such as hunting and fghting, is to train the young in imaginative thinking
for use in adult activities. Again, we see attempts to explain creativity in the
context of mammals. If we could study creativity in simpler animals, even
such as insects, perhaps the pursuit would become easier.
Despite simple nervous systems, insects apply different strategies to identify
landmarks, detect objects, and plan courses of action. They are not mere
refex machines, but use memory, evaluation, and perceptual mechanisms,
including the detection of geometric shapes and route-segmentation, to
function in their environment depending on previous experience (Collet &
Collet, 2002; Giurfa & Menzel, 1997). Some might fnd it diffcult to envi-
sion that creativity has anything to do with insects and as such there have
been no attempts to shed light on creativity in terms of simpler animals than
mammals. However, exploring the general concept of cognition in animals
lower in the phylogenetic tree has gained some attention in recent years (van
Duijn, Keijzer, & Franken, 2004). Godfrey-Smith (2001) recently proposed
that cognition evolved to cope with environmental complexity and Maturana
and Varela (1973) introduced theories of the origins of cognition with the
concept of autopoiesis; a network of component-processes that perpetually
generate themselves. Autopoiesis provides support for theory proposed in
this chapter, and is discussed in more detail later.
Exploring cognitive mechanisms in animals, especially insects, is hard because
they cannot tell us what they are thinking. Complex processes of intelligence
cannot be readily explored in natural animals and neither can the evolution
of the mechanisms involved. The methods offered by artifcial intelligence
present optional approaches to gaining a detailed understanding of intel-
ligence (Thrisson, 2004). An added beneft of such research is (sometimes
unexpected) practical uses of the resulting technologies for various purposes.
But creating simulated environments and organisms with suffcient realism to
be considered valid is no simple task. The environment and organisms must
capture realistic features essential to the specifc questions studied without
excluding possible contributing factors, which are present in the correspond-
ing natural systems.
Thrsson
Cellular automata (CA) are discrete, dynamical systems, which have been
used frequently in research on artifcial life (ALife), in which the emergent
organization of the cells themselves result in agent-like behavior. The works
of Antony, Salzberg, and Sayama (1998, 2004) are good examples of these
kinds of artifcial life simulations. The lifeforms which emerge are primi-
tiveallowing a very limited spectrum of association with higher functioning
organisms. This is because most approaches limit the whole system to a CA
framework, explicit rules governing everything. In CA-based ALife research,
the term agent generally used to refer to a cell encompassing complex
internal processing. The interactions of such cells constitute multi-agent
simulations. The model presented in this chapter must not be confused with
such ALife experiments, as it consists of preprogrammed agents modeled
after natural insects, and uses cellular automata only to produce landscape
with emergent properties. While CA-based ALife simulations are mainly
concerned with investigating the emergent behavior of interacting agents, we
are exploring notions of how emergent environments affect creative aspects
of intelligence on onto- and phylogenetic levels.
A.Simulation.Approach.to.Analyzing.Creativity
The primary hypotheses of interest here are that the evolution of creativity
mechanisms are directly related to the complexity/diversity of emergent
environments, and that the underlying mechanisms of creative behavior are
essentially the same in all creatures, despite the obvious human advantages
mentioned above. We propose the theory that creativity is evolutions answer
to perceptually apparent unpredictability. This is different from some hypo-
thetical or actual unpredictabilitythe discussion here revolves around
unpredictability from the standpoint of the creature.
H
1
: An increase in environmental diversity will result in a larger set of
distinct behavioral patterns (plans).
H
2
: If the world is too complex, there will be no persistent structure evident
in the composition of plans. Creative capabilities of all creatures are
bound by rules inherent in the environment where they evolve.
More precisely, the previous hypotheses depend on (a) the number of indi-
vidual components in the environment and their interactions, (b) the creatures
ability to perceive their environment, and (c) the creatures set of operators
which can be applied to modify their behavioral pattern (in particular, to
make plans).
Creatures inhabiting a simple, closed, and static world would tend to evolve
to become completely robotic due to predictability of the environment; if an
environment is very simple the cognitive system of the creatures require little
effort to survive. In a complex world, where events in the environment are
not evident entirely by observation of the current situation, evolution must
provide organisms with a mechanism to predict in uncertain situations. With
the increase of interacting components in the environment, more complex
cognitive efforts are thus required to produce and assess reactions to the
current situation.
Earths environment is governed by rules that can be traced down to a sub-
atomic level, resulting in its emerging attributes such as land and sea, water
circulation, and ecosystems. As creatures evolve in worlds governed by these
rules, their plan-making mechanisms come to refect them. If a particular world
should on the other hand be random-based, a creatures planning ability might
produce more diversity, but there would be no logic as to when or where the
plans are applied. This makes it important to measure the complexity of the
world in addition to the diversity of the creatures.
We propose to measure the structural complexity by the use of cellular au-
tomata where, through simple interacting rules, complex overall behavior
emerges (see World). The fact that there are general rules in the environment,
many of which are independent of (perceived) scale, indirectly lends support
for the proposed hypotheses.
Creativitys roots, according to this line of thought, derive from interactions
between the environmental structures: If the interactions of these structures are
not obvious (to the organism), perceivable, and very repetitive, creatures must
evolve an ability to internally represent and link unobvious causationsthus
providing them with greater capabilities to produce diverse behavior in response
to the diversity of the environment. Another reason for such a requirement
is that visual cues may appear once it is too late for the creatures to react,
supporting the evolution of a mechanism capable of predicting possible sce-
narios before they actually occur (imagination). Cognitive mechanisms must
Thrsson
therefore be adept at predicting the environmentand in parallel keeping
plans created for survival logical and relative to those predictions.
We will now describe a framework and an experiment created for exploring
the hypotheses. Then we will explore their deeper theoretical support.
Simulation.Framework
In the present framework the creatures are still at the evolutionary stage of
relative reactivenessthey depend solely on visual cues in the environment
to initiate a plan. However, the plans themselves directly serve as a prediction
of the environment and are suffcient to produce results which support or
refute the main hypotheses. The framework rests on the theory that creativity
is a product of certain environmental characteristics that manifest themselves
in organisms as they evolve in that environment. It purports that creativity is
evolutions answer to a perceptually apparent unpredictability of emergent
environments (H
1
), and that creativity and logic co-evolve as imposed by
the causal-relationship between levels of organization in emergent structures
(H
2
). To explore these notions through means of computer simulations, we
have devised a program for simulated worlds that exhibit emergent proper-
ties, and creatures that are able to sense and act in these worlds. The world is
composed of cellular automata, the creatures of genetically-controlled cogni-
tive faculties (Thrisson, 2004). Each individual creature learns and exhibits
intelligent behavior during its lifespan with the use of primitive associative
memory, pattern recognition, and planning. Thus, evolution does not effect
the creation of plans directly, but modifes the wiring of the creatures learning
mechanism, who in turn dictates which methods are used to generate plans.
Let us emphasize that evolution is perturbing a single mental model
1
, not
constructing new ones. This is important as were not attempting to evolve
new mental components but observing how evolution makes use of readily
available components, depending on different environmental contexts. The
simulation is turn-based (updates once for every 50 steps of the creature).
The creatures are run one at a time to simplify the simulation
2
.
As we argue that creativity is an aspect of intelligence in general (Boden,
2004), in context with the environment in which the intelligence operates, it
can not be measured explicitly at this point. Rather, we observe what types
of behavior the creatures exhibit, how they contribute to our perception of
creativity, and what kind of mechanisms are used to produce them. The
observation provides support for our explanation and new defnition of cre-
ativity, independent of any particular cognitive mechanism (see theoretical
discussion).
The World
Cellular automata (cf. Wolfram, 2002) have frequently been used in artifcial
life (ALife) research where the cells are programmed to organize themselves
to form patterns which behave like creatures (Antony et. al, 1998). The world
in our framework is a classical cellular automaton with cells made to represent
physical phenomena such as rocks, earth, and grass (see screenshots later
in this section). This approach differs from other work on cellular automata
in that the creatures that inhabit the world are separate entities with prepro-
grammed cognitive systems (they are not cells), and interact with the worlds
cells via perception and motor abilities.
Cellular automata was chosen because of its similarities with natural envi-
ronment emergence, and because of the ease of quantifying environmental
complexity. As discussed earlier, the Earths environment is composed of
emergent structures produced by the interactions of its constituent subcompo-
nents. This effect is captured by cellular automata as the interactions of cells
results in emergent structures in a similar manner. Very complex behavior
can be produced by the use of simple rules such as those used in this experi-
ment while at the same time keeping the world logical in terms of general
behavior since the state of an individual cell is dependent on its own state
and/or surrounding neighbors.
The cellular automata worlds in this experiment are 2D, 100x100 squares
wrapped around to form an infnite plane with a repeating pattern; if a creature
walks off the edge of the screen, it enters the world again on the opposing
side. In the complex dynamic world used in this experiment, the cells in the
simulation are dependent upon 25 of their surrounding cells. Each time the
world updates, every one of the 10,000 cells looks at the 25 surrounding
neighbors and calculate their next state according to given rulesresulting
in 250,000 cell checks each time the world turns. The world is set to step
at 1/50 steps of the creatures. By varying the rules and number of different
kinds of cells, the diversity of the environment can easily be measured since
the variation of cells and the number of their rules constitute the worlds
complexity/diversity. The rules are not sensitive to neighbor positions, only
Thrsson
quantity and lifetime. Adding rules to the components of the world increases
the dynamism of the simulation, but this must not be confused with increased
speed or, alternately, less time to think for the creatures: The simulation is
turn-based (updates once for every 50 steps of the creature), and as such,
increasing dynamism of the environment only results in the environmental
structures becoming more varied and complex. Cognition is always allowed
to run its full course before the next turn.
The worlds complexity is quantifed by counting the number of component
types (cell types) and the number of rules these components abide. Complex-
ity measurement is defned as:
Qc = Pi * C
where C represents the number of component types perceivable by the creatures
and Pi represents the number of rules pertaining to these cells. A rule is defned
as any rule or set of rules that causes a state transition for a cell under some
condition or conditions. This can be taken as an approximate estimate of the
complexity of a world because each rule represents a potential for structural
change. While this method can certainly be improved upon, it provides us
with an indication of the potential complexity of the environment.
Creatures
The artifcial species should be designed to bear a resemblance to natural
insects found on earth, this information is based on the research of Collet
et al. (2002). The simulation of insect memory and navigational abilities
has been tried successfully using a physical LEGO robot (Chan & Wyeth,
1999) and other physical robots (Mller, 2000). The methods are similar to
those employed in this framework: Insects use image matching to associate
visual cues with actions to navigate to and from nest and food sources
3
. The
method for associative learning and cognitive apparatus used here is a sim-
plifed version of the one applied in the LEGO robot (see perception cortex
and episodic memory).
We assume that the physical properties of our artifcial cognitive system
have already evolved to an insect-like state, having primitive memory and
perceptual abilities, along with a system for the creature to be aware of
the consequences of its actions (loss or gain of energy after execution). The
creatures cognitive system is divided into several parts, their physical body
comprising a (a) vision system, (b) mobility, (c) a mouth for feeding, and
fnally a (d) mental apparatus for interpreting visual information and generate
plans. The cognitive systems components are depicted in Figure 2. The visual
information allows them to form memories, which then serves as the basis of
forming repeatable plans that enable them to interact with their environment.
The plans are composed of primitive actions, such as go-forward, turn-left,
turn-right, and eat (see details in Plan Composer). The creatures have a single
goal: To survive as long as possible by fnding, recognizing and consuming
food and passing their genes on to their offspring. In our simulation, green
cells (grass) provide the creature with energy if eaten. Eating brown cells has
no effect, and eating gray cells causes a reduction of energy. Only the fttest
creatures, those that survive the longest in each generation, are allowed to
reproduce. Inheritance of cognitive mechanisms is achieved using simple
genetic algorithms. The genes (see Control Network) of a new individual are
mutated at birth to produce diversity in each successive generation.
We will now continue with a detailed description of each part of the artifcial
creatures cognitive system.
Perception Cortex
The perception cortex deciphers visual information in the form of quantities
of colors and registers which part of the visual feld a dominant color is if

Figure 2. The perception-action-loop of the agents created in the project
(Numbers trace the order of events)
Thrsson
there is any color more evident than others. The visual feld of the creature is
rectangular; tuned to perceive a 7x7 grid of cells. When the cortex receives
input it begins interpretation by registering the number of each distinct colors
it perceives. The visual feld is split into four zones, each one marked for its
dominant color. The zone which contains the most dominant color overall
is marked as the dominant zone. Number of colors, dominant zone, and the
dominant color serve as input to the control network (see section on Control
Network) during plan creation.
Decision Cortex
The decision repertoire consists of mediating information between all the
cortices, keeping a track of current processes, and initiating new processes.
The cortex shares the responsibility of creating a new plan with the control
network (see below), whether to modify old plans or to create a new one
completely from scratch.
During each discrete step in the simulation the decision cortex assesses
the status of the creature and its perceptual input and decides which action
should be taken. While there is a plan is being executed, each primitive action
executed by the motor cortex is routed to memory (plans are sequences of
primitive actions), the score of the action is calculated and associated with the
current perceptual input. The score of a plan or primitive action is determined
by the energy loss/gain before and after its execution. The second mode is
if there is no current plan executing: If a plan execution was fnished in the
last step, its total score is calculated and sent to memory. Next a request is
sent to the episodic memory for a recollection of similar situations (see the
next section) and to the control network to create a new plan depending on
the situation. To sum up, the decision cortex is responsible for
a. Calculating the score of a plan that just fnished executing.
b. Requesting a memory search in episodic memory.
c. Routing the perceptual input and physical actions to memory for stor-
age.
d. Requesting a new plan from plan composer.
e. Sending new proposed plans or actions to the motor cortex.
Episodic Memory
The episodic memory is the storage area for all perceptual input, the crea-
tures actions, and the consequences of those actions. Memories include a
particular visual feld, the plan that was associated with the perceptual input,
the scorein particular the energy expenditure, which was registered during
the execution of the plan, along with a normalized plan score.
During each step, the decision cortex sends a message to the episodic memory
to fnd all memories that the creature might have of similar situations. The
procedure is as follows:
Find all similar memories by frst rating them along three semi-independent
feature dimensions:
1. Image.Matching.(dimension.one): The visual feld in general: Com-
pare each cell of the current visual feld to each patch of a particular
memory M. Similarity scoring is incremented with each cell that cor-
responds precisely to a cell in the same location in memory M relative
to the creature (thus, if a cell in the upper left-hand corner is green, and
the upper-left hand cell in memory M is green, the memory gets a point
for similarity). This score is normalized, so the increment for each cell
thats identical is 1 divided by the total number of cell in the visual feld.
The fnal score is the sum of all identical cells.
2. Dominant. Zone. Similarity. (dimension. two): Compare the current
visual felds dominant color and the zone the dominant color is mostly
evident in to all memories. The score is Boolean.
3. Patch.in.Front.of.Mouth.(dimension.three): Compare the patch in
front of the creatures mouth to that same patch in all existing memo-
ries. This is similar to the zone similarity scoringif the current patch
in front of the mouth corresponds to that same patch in a memory, the
similarity is turned on (again, 1.0 or 0.0).
For each of these features, the memories are arranged according to each
features similarity score. The plan associated with each memory (there is
either a plan or a primitive action associated with each memory) is retrieved
for the top three memoriesone for each dimension. Plan scores (energy
payoff) are normalized by the following equation:
0 Thrsson
S
f
=1 / S
p
(S
max
S
min
)
where S
f
is fnal score, S
p
is the original plan score, S
max
is the maximum
plan score of any plan retrieved from memory, and S
min
is the minimum
score for any plan retrieved from memory. The normalized plan scores of
memories in the sorted dimensional arrays are multiplied by the similarity
measurement of the memories, providing a winner with regards to similar-
ity and plan score.
Control Network
The control network is the part of the program that gets passed on and mu-
tated between generations; it holds the creature genome, which constitutes a
set of control boxes and the connections between them. The control network
is made up of 70 boxes, which consist of two inputs I
1
and I
2
, an output O
and an operator, which determines what to do with the inputs received. The
types of boxes in the control network are:
Static. boxes have a fxed output that stays constant throughout the
lifetime of an individual.
NOR boxes.or Not OR boxes return 1 if both inputs are zero.
NAND.boxes or Not AND boxes always return 1 except if both inputs
receive numbers.
Input.gate.boxes return I
1
as output only if the value of I
2
is more than
zero.
Each of the boxes serves as a node in the control network. Outputs from the
boxes are connected to none, some, or all of the other box inputs. The exact
routing of the control box inputs and outputs is determined entirely during
reproduction of the creatures. That is, the creatures genetic confguration is
responsible for using these connections to activate plan composition mecha-
nisms. There are four instances of each kind of box in a control network
except for static boxes, which are 16. For each box, a total of four possible
connection confgurations are possible. Each of the 30 boxes can receive two
inputs. There are no initial connections between control boxes for the frst
generation in all worlds, and the initial outputs of static boxes are set to zero.
After the frst generation, all settings of control boxes and their connections
are changed by mutation. Following the terminology introduced by Holland
(1998), which covers natural and artifcial adaptive systems, we can describe
the adaptive process of the control network.
The powerset of control box confgurations and connections in the network is
denoted . We denote a specifc control box confguration A as a set of pos-
sible confgurations of connections between them: A ={a
i1
, a
i2
, ... a
iki
}.
The combinations of control boxes and their connections are defned by the
adaptive plan which uses the mutation operator between generations. The
criterion for comparison of is the age a creature reaches by using plan-
making mechanisms (which are in turn controlled by the particular adaptive
plan). The set of structures (control network confgurations) attainable by
the adaptive plan is represented by the following equation:
= A
1
A
2
A
n
=
n
i=1
A
i
Since the environment, E, selects over time, which control boxes survive
between generations, an evolving, adaptive plan is defned over time by a
particular environment E. This is depicted in Figure 3; the big box represents

Figure3.AdaptiveprocessesafterHolland(1998).A(n)denotesaspecifc
confgurationofthecontrolnetwork.
Thrsson
which contains all structures attainable by the adaptive plan , which is
decided by the environment E using . The adaptive plan produces a trajec-
tory through , which are specifc confgurations of the connections between
control boxes, connections to the plan composer (see next section) and the
output of the static boxes (a number between 1 and 30).
On every step during execution, the control network receives a request from
the decision cortex for updating its outputs. The outputs feed into the plan
composer and set control parameters as the creature decides how to plan its
actions. There are three main ways for planning the next step, as explained
in the next section.
Plan Composer
The plan composer creates plans from sequences of primitive actions (go-for-
ward, turn-left, turn-right, do-nothing and eat). During each simulation step,
the plan composer receives instructions from the control network regarding
what to do and specifcally how to do it. The instructions are in the form of
integers from the control network inputs into the plan composer. They are
deciphered by simple logic gates. The methods that the plan composer can
use to create the next plan are the three methods:
a. Create a new plan from scratch.
b. Combine halves of two old plans.
c. Mutate an old plan (randomly change primitive actions of an older
plan).
4
During plan-making, the creatures can either create a new plan from scratch
or use one of the other two options. When creating a new plan from scratch,
the plan composer randomly selects the length of the plan and how many
instances of each primitive action it should include. With random distributions
of actions, the other two methods, combine halves and mutate, become very
important as they provide a much more controllable way of making sensible
plans. Also, by using combine, mutate and using an old plan unmodifed, the
development of an individual will become more evident as the creature is
bound to use the methods on plans that have provided good results (memo-
ries are organized by score, see Episodic Memory). Note that in order to use
methods b and c, the creature has to have created at least one random plan
from which it can build a modifed plan.
The use of a random mechanism to produce new plans is a matter of conve-
nience and deserves some discussion. It is not our claim that creativity is a
random process. On the contrary, we propose that creativity is an aspect of
intelligence, independent of any particular cognitive mechanism. The main
concern is to have a mechanism, which generates new plans from scratch,
using other means than recombinations of previous plans. This allows us
to compare and contrast the evolutionary incentive to produce completely
new plans (regardless of the method of production) as opposed to relying on
experience. Exactly how such novelty-producing mental components work
on higher levels of natural intelligence, such as the human level, is another
matter and not the focus of this chapter.
Experiment.
In the experiment, we use three versions of the world. Simple (Es), which
is made up of stripes alternating green and gray (Figure 4). Green is grass
(food), gray is rock, and brown is mud (if the creature eats a green cell, the
cell turns brown). Es has Qc =3. Simple Dynamic (Esd) is the same as Es
with the seemingly small change that the stripes change regularly from gray
Figure4.Ascreenshotofasimple(Qc=3)andsimpledynamicworld(Qc
=9);100x100cells
Thrsson
to green and green to gray, increasing the worlds Qc to 9. Update frequency
of Esd is every 50 steps. Both Es and Esd reset for each new individual. The
third world, Complex Dynamic (Ecd) presents more landscape-like structures;
Qc =18 in Ecd (Figure 5). Ecd begins with initial random placement of cells,
but is updated 500 times to eliminate the initial randomness before creatures
are introduced to the world.
Each generation is made up of 100 creatures (Ng =100), natural selection is
implemented by sorting the creatures according to age; the top ten individuals
(C
1
, C
2
C
10
) are selected as the basis for the next generation. The selec-
tion criteria for inheritance, , is the creatures age reached. The creatures
only goal is to survive. Having a single goal simplifes comparison between
the different worlds and plans. Because the creatures only have one goal, all
plans that the creature creates are geared towards reaching that goal. And thus,
coming up with alternative plans for reaching a goal is a form of creativity.
Results.and.Interpretation
The results show a high similarity between simple (Es) and simple dynamic
(Esd) environments in the creatures planning behavior. In both environments,
Figure 5. A screenshot of a complex dynamic world, Qc = 18; 100x100
cells
the creatures used all of their available mechanisms for plan creation (create-
new-plan from scratch, mutate-old-plan, and combine-old-plans). Figures 6
and 7 show the average use of the three mechanisms for plan creation over
the lifetime of the creatures in each of the 100 generations (see section on
creatures for details about the mechanisms). While the creatures are producing
new plans from scratch relatively often, they compensate for plan randomness
by using the other mechanisms. The creatures evolving within these simple
worlds are maintaining a balance between innovation and logic: creating new
and untried plans as well as learning from their previous experiences by using
and combining plans that have proven effcient and useful.
In terms of the proposed hypotheses, it would not have been surprising to see
more difference in the results of the two worlds. Looking at the two simple
worlds only, the main hypothesis, that an increase in environmental diversity
results in a larger set of plans, is not clearly supported or disproved. The
similar results might be due to structural likenesses between the two worlds.
When compared to the results from the complex dynamic world (Ecd), how-
ever, the hypothesis is clearly supported. Figure 8 shows the average use of

Figure 6. Es. average use of create new plan (circles), mutate old plan (boxes)
andcombineoldplans(triangles)X-Axis:generation10-100;Y-Axis:0-30
Thrsson

Figure 7. Esd. average use of create-new-plan (circles), Mutate-old-plan
(boxes) and combine-old-plans (triangles) X-Axis: generation 10-100; Y-
Axis:0-25

Figure 8. Ecd. average use of create-new-plan (circles), mutate-old-plan
(boxes)andcombine-old-plans(triangles)X-Axis:generations10-100;Y-
Axis:0-20
the plan-making mechanisms over the 100 generations. With Qc =18, the
creatures resorted to using the creation mechanism (creating a new plan from
scratch every turn), while leaving the other mechanisms relatively unused.
The plans and behavior of the creatures was therefore very diverse but the
lack of learning from previous experiences resulted in their plans maintaining
no logicality
5
. The creatures had shorter lifespans in more complex worlds
and could therefore not use the plan-making mechanisms as often. In Es the
creatures average age reached is 28; in Esd it is 25 and in Ecd it is 21.
Figure 9 shows an example of how the creatures plan maintains logicality
by eating only the food (green) while skipping the rock-lines (gray) which,
if eaten, result in a substantial loss of energy. Similarly, Figure 10 shows a
creature having turned to eat a row of food in a consecutive manner. Such
behavior would not have been possible were it not for the fact that the crea-
tures in Es and Esd were using the mechanisms that build on previous expe-
rience. In the Ecd (Complex Dynamic Environment), the creatures showed
little logicality in their behavior due to their frequent usage of the random
plan-making mechanism (Figure 11 shows a creature within the Ecd). To
fnd food in Ecd is not as simple as fnding food in Es and Esdthe agents
have to produce behavior that allows them to cope with the complexity of
their environment. The agents work relatively mechanically in Es and Esd

Figure 9. A creature (black) in an Es world, eating only the food (light gray),
bypassing the rocks
Thrsson
because the visual cues are few and provide accurate accounts of what will
happen if the agents execute a certain plan. If the agents encounter a gray
cell-structure (a stripe) in Es, they can depend on the fact that if they use the
plan move forward once, then eat they will gain energy, as the cell they will
eat is guaranteed to be green (food). However, in Ecd, the green structures
behavior is much more diverse, and hence harder to predict. This difference
of environments are comparable to those of bacteria and insects in the sense
that insects partain to environments of much higher diversity.
In summary, the results strongly suggest that creativity is governed by en-
vironmental complexity, and that structural coherence dictates logicality
of produced plans. Particularly, agents whose cognitive system evolve in

Figure10.Acreature(black)inEseatingarowoffood(lightgray)consecu-
tively.Theimagehasbeenenlarged.Thescreenshotshows10x6cells.
Figure 11. A creature (black) in Ecd lurks behind a gray rock formation, away
fromthefood(lightgray).Thescreenshotshows31x17cells.
overly complex worlds tend to create diverse (random) plans, but simple
ones impose plans that maintain novelty and diversity without venturing into
randomnesssupporting the stated hypothesis.
Theoretical.Discussion
We argue that a coherent explanation of the different processes and mecha-
nisms of creativity must be done in terms of self-organization and emergence.
As creativity relies on many cognitive functions, we must incorporate those
functions and discuss their origins.
Cognition is rarely associated with the behavior of reactive organisms. As
noted by van Duijn et al. (2006), anthropocentric interpretations of cognition
introduce a dichotomy between reactive/infexible organisms and refective
cognizers. An anthropocentric view has also been prevalent throughout re-
search on creativity. Alternative theories to popular views of cognition have
been proposed, explaining cognition from a simpler and more mechanical
viewpoint. In 1973, biologists Varela and Maturana introduced the concept of
autopoiesis as a redefnition of life. Their explanation is especially relevant
to the concept of emergence and complex systems as it presents a bridge
between structurally automatic systems and cognitive systems. In the words
of Varela and Maturana (1973), an autopoietic system is:
...amachineorganized(defnedasaunity)asanetworkofprocessesof
production (transformation and destruction) of components which: (i) through
their interactions and transformations continuously regenerate and realize the
network of processes (relations) that produced them; and (ii) constitute it (the
machine) as a concrete unity in space in which they (the components) exist
by specifying the topological domain of its realization as such a network.
(p. 78).
A popular example of an autopoietic system is a biological cell. The cells
components are molecules, interactions are chemical reactions and the cell
membrane the physical boundary that forces reactions into an entity distin-
guishable from the external environment (Beer, 2004). Computer simula-
tions of autopoietic cellular structures have been created where proto-cells
emerge from a digital primordial soup, gradually gaining in complexity to
0 Thrsson
form membranes, in many ways similar to those observed in natural systems
(Madina, Ono, & Ikegami, 2003).
Although the concept of autopoiesis remains somewhat controversial (Beer,
2004; van Dujin et al., 2006), it can help us explain the origins of creativity
in a unifying manner. While admittedly there is a gigantic leap from auto-
poietic systems such as a biological cell to an insect-like agent such as those
used here, the methods that organisms (or autopoietic systems) can use to
maintain their structure can be generalized to the identifcationofapotential
threat to its organization and responding appropriately; a perception-action
coupling. van Duijn et al. (2006) have claimed in relation to autopoiesis, that
minimal cognition is a dynamic and adaptive sensorimotor process, and have
noted that although bacteria have no nervous system, molecular sensorimo-
tor mechanisms of most bacteria closely resemble some of the functional
sensorimotor features of nervous systems. This allows us to generalize the
cognitive functions of organisms at different levels of complexity as different
means of maintaining the perception-action coupling
6
.
The following is a proposal for a new defnition of creativity and a framework
for how the different concepts of cognition co-evolve with the phenomena.
We will then proceed to explain how these concepts are interralated in terms
of emergence and autopoiesis.
1. Creativity is an agents ability to vary and adapt processes in order
to fulfll its goals. Processes can be internal (mental functions), or an
externalization of internal processes (physical actions and interactions
with the environment), where an agent is a living (autopoietic) structure,
be it organic or artifcial. Creativity can be exhibited without explicit
understanding of the environment or goals, and the diversity of prod-
ucts or artifacts produced through creative processes are limited by the
physical attributes of the agent. Creativity is an aspect of intelligence as
a holistic system, independent of any particular cognitive mechanism.
2. Logicality is an organisms ontogenetic ability to produce actions which
are coherent with the organisms goals and structural organization of the
environment. Environmental organization imposes itself on organisms
as they evolve and is refected in logical behavior.
3. Insanity is the inverse of logicality, resulting in behavior which is not
applicable to the context in which they are realized, failing to fulfll the
agents goals and are consequently illogical or nonsensual.
4. Imagination is an organisms mechanism for modelling its environment
internally to predict a turn of events. When creatures gain in organizational
complexity along with the environment, reactive mechanisms become
insuffcient to predict the environment. This becomes an impetus for
other, more complex mechansisms to evolve. In particular the ability
to represent the environment as a mental simulation.
The previous defnitions include several concepts of intelligence and cogni-
tion, mainly: (a) perception and action, (b) goals, (c) logic and insanity, (d)
understanding, and fnally (e) imagination or internal representations. We
will now proceed to explain these concepts and their relationships in more
detail.
To explain how anorganismcanfulfllgoals, we must establish what goals
are. It can be said that autopoietic systems preservation of structure is an
implicit goal, realized by the environment-agent coupling. The interactions
that maintain an organisms organization, thereby allowing it to maintain
identity as one, are conditions internal mechanisms of the organism aim to
preserve. It is more than possible that this basic principle is the foundation
from which continued complexity of hierarchical organization can emerge.
Gradually, these systems organize a multileveled hierarchy of processes which
constitute what we perceive as, for example, insects. Goals are consequently
always evident as particular conditions of the environment. At higher levels
of complexity (insect level, for example), organisms seek out these condi-
tions by producing planswhether they are instinctive or refective. In other
words, goals are conditions which allow an organism to persist
7
. Intention-
ality, in this sense, becomes the intellectual process of understanding these
conditions and how they relate to the agent.
To explain logics relationship with creativity, we must explore how envi-
ronmental complexity affects goal-oriented behavior. The environment is
constituted by emergent structures of varying complexity; the agents depend
on visual cues from this environment to produce actions in order to preserve
their structure. In our simulation, the agents predict the environment by
generating random actions to begin with, and, over a lifetime learn to associ-
ate the correct actions to each perceptual stimuli. When the world becomes
more complex, it has two major implications for the agents: The diversity of
perceptual stimuli greatly increases, meaning that the domain of plans with
which the agents can associate perceptual cues increases. Secondly, as the
environment is not highly repetitive, the agents must produce behavior to meet
Thrsson
this diversity; different plans apply to different problems (or multiple plans
can apply to the same problem). This was clearly demonstrated in the results
of our experiment. With respect to this relationship between the environment
and plans, it can be stated that the plans are refecting the architecture of the
environment. Over time, evolution provides the agents with mechanisms to
produce plans, without direct control over the plans themselves. Over the
lifetime of individual agents, these mechanisms are fne-tuned to match the
specifc circumstance and refect the environment more accurately. If the
plans are not coherent with the specifc circumstance they are presented in,
they are illogical or irrational; they fail to comply with the agents goals. It
can consequently be stated that logic is bound in the coherence of the plans
(behavior) and the goals (or environmental conditions) that the plans are
intended to fulfll.
At this point weve established that plans are predictions, as well as refec-
tions, of environmental organization. Even though organisms are not very
advanced, the correlation between environmental organization, goals, and
behavior imply that the organisms cognitive system has an (implicit) un-
derstanding of the environment. In our experiment, however, they have no
internal mechanism to understand, which presents an obvious evolutionary
disadvantage; without an internal representation, the cause for a complex
emergent structures behavior is perceptually invisible and unpredictable to the
organism. In other words, the cognitive mechanism has to experimentally try
out different kinds of plans in order to become more adept at fulflling goals.
An internal representation of an emergent structures constituent components
and their interactions would be needed to accurately predict its behavior
8
.
From an evolutionary perspective it can be assumed that, with increased
complexity, natural organisms evolve more advanced methods of prediction,
allowing them to represent structures internally and simulate or imagine
environmental events. An internal representation of an emergent structure
enables a deeper understanding of that structure, and hence, a more accurate
prediction and increased chance of survival. Understanding can be equated
with that of explanation (Baas & Emmeche, 1997). Using the simulation
of this chapter as an example, we can explain that a brown cell turns green
because it has a certain number of green neighbors surrounding it. Moving
up a level of complexity, we can explain the group behavior of the cells by
reference to the function of each individual cell, and so forth. If the agents in
our simulation would be capable of understanding why (or in what context)
the food appears (green cells)they could predict the overall behavior of
the structures that emerge (and decrease the number of random plan genera-
tions). Instead of wasting energy following the food, for example, they could
intentionally select a spot to sit down and wait for the arm of a spiral to come
to them (a creative solution).
Following this line of thought, it is more than likely that mechanisms of
internal representations adopt the emergent properties of the environment.
Imagination, humans most valued creative posession, is possibly an advanced
simulator of emergent phenomena. Such spontaneous generation of new (and
surprising) ideas, from components or ideas that otherwise did not appear
to present any novelty, has been noted in research as similarities between
creativity and emergence (Gero, 1996; Liu, 1996).
Future.Trends
Advances in computer science are enabling simulations of increasingly
complex systems. With such powerful tools at hand, it is only natural that we
continue developing methods for exploring the potential origins and nature of
intelligence. There are many aspects of creativity which remain unexplored;
the premise of this chapter is creativity in its most primitive forms, leaving
out questions of how mechanisms of more advanced organisms contribute
to, and affect overall, creative abilities.
Explaining and exploring cognition in terms of simpler organisms is gain-
ing support in the scientifc community, and our experiments suggest that
this might provide us with a better foundation for explaining high-level
intelligence as complex systems. We believe the same applies to creativity.
The emphasis should be on exploring how more advanced mechanisms of
cognition are related to creativity, and how such relationships are brought
about by evolution.
It is important to start attempting the unifcation of various aspects of intel-
ligence; we have spent the last decades dissecting and separating the mecha-
nisms involved and now it is time to put the pieces back together and explain
how they work in unison. Complex systems simulations are an imperative
for studying how creative functions of organisms are interrelated and, should
emergence prove as crucial to this explanation as we propose, they are close
to indispensable.
Thrsson
Conclusion
We presented an explanation for the hypotheses that creativity is evolutions
answer to environmental complexity and that logic and creativity co-evolve
as dictated by emergent, structural coherence. Using emergence and autopoi-
eses as a foundation, we defned creativitypresenting a unifying framework
for the related cognitive concepts of perception and action, goals, logic and
insanity, understanding and imagination. The theoretical arguments are in
their initial stages, but offer much potential for further investigation and
experimentation, especially in terms of computer simulations.
Relatively complex creatures are presented in the chapters described experi-
ment, which evolve within worlds with emergent environmental properties.
The creatures were modeled after natural insects, utilizing essential cognitive
mechanisms such as visual perception, memory and planning mechanisms,
which was important for explaining and defning the concept of creativity.
Cellular automata proved useful for simulating environmental emergence
and quantifying the complexity of emergent worlds.
The artifcial creatures in the experiment showed increased diversity of
behavior in overly complex worlds, maintaining little or no logicality as to
when plans were applied. The results support our hypotheses, suggesting
that environmental complexity and organization affects the evolution of
creativity to a high degree.
Future work using the simulation framework in this chapter will involve im-
provements of the creatures cognitive system, and (a) increasing the gene pool
and number of generations in each simulation run (it is not uncommon that
similar experiments run thousands of generations), (b) trying more types of
worlds of different complexity levels. We also note that the custom cognitive
architecture used here might possibly be too limited to allow the creatures
to adapt to worlds of high complexity, and that more robust solutions, such
as artifcial neural networks, might offer a better solution.
Acknowledgment
I would like to thank the following individuals for their advice and continued
support: Dr. Kristinn R. Thrisson, Dr. Hannes Hgni Vilhjlmsson, and Dr.
Yngvi Bjrnsson of Reykjavk Universitys artifcial intelligence laboratory,
the Center for Analysis and Design of Intelligent Agents.
References
Adams J . (1980). Conceptual blockbusting. New York: Norton.
Antony, A., Salzberg, C., & Sayama, H. (1998). A closer look at the evo-
lutionary dynamics of self-reproducing cellular automata. Selection,
J une, Volume 2.
Baas, N. A., & Emmeche, C. (1997). On emergence and explanation. Intel-
lectica, (25), 67-83. Retrieved J une 28, 2004, from http://www.utc.
fr/arco/publications/ intellectica/n25/25_04_Baas.pdf
Beer, R. D. (2004). Autopoiesis and cognition in the game of life. Artifcial
Life, 10, 309-326.
Boden, M. A. (2004). The creative mind: Myths and mechanisms. New York:
Routledge.
Boden, M. A. (1995). Creativity and unpredictability. Stanford Education
andHumanitiesReview,4(2).
Carruthers, P. (2002). Human creativity: Its evolution, its cognitive basis,
and its connections with childhood pretence. British Journal for the
Philosophy of Science, 53. Retrieved May 19, 2004, from http://www.
philosophy.umd.edu/people/faculty/pcarruthers/Creative-thinking.htm
Chan, P., & Wyeth, G. (1999). Self-learning visual path recognition. Pro-
ceedings of the Australian Conference on Robotics and Automation
(pp. 44-49), Brisbane.
Collet, T. S., & Collet, M. (2002). Memory use in insect visual navigation.
Nature Review, 3, J uly.
Gabora, L. (2005). Creative thought as a non-Darwinian evolutionary process.
Journal of Creative Behavior, 39(4), 65-87.
Gero, J . S. (1996). Creativity, emergence, and evolution in design: Concepts
and framework. Knowledge-Based Systems, 9(7), 435-448.
Giurfa, M., & Menzel, R. (1997). Insect visual perception: Complex abili-
ties of simple nervous systems. Current Opinion in Neurobiology, 4(7),
505-513.
Thrsson
Godfrey-Smith, P. (2001). Environmental complexity and the evolution of
cognition. In R. Sternberg & J . Kaufman (Eds.), The evolution of intel-
ligence (pp. 233-249). London: Lawrence Elrbaum Associates.
Holland, J. H. (1998). Adaptation in natural and artifcial systems. Cambridge,
MA: MIT Press.
Liu, Y. (1996). What and Where is design creativity: A cognitive model
for the emergence of creative design. Paper presented at the IDATER
1996 Conference, Loughborough, UK.
Madina, D., Ono, N., & Ikegami, T. (2003). Cellular evolution in a 3D lattice
artifcial chemistry. In W. Banzhaf et al. (Eds.), Paper presented in the
proceedingsofECAL03(pp. 59-68). Dortmund: Springer.
Martindale, C., & Greenough, J . (1973). The differential effect of increased
arousal on creative and intellectual performance. J ournal of Genetic
Psychology, 123, 329-35.
Maturana, H. R., & Varela, F. J . (1973). Autopoiesis: The organization of the
living. In H. R. Maturana & F. J . Varela (Eds.), Autopoiesis and cogni-
tion (pp. 59-138). Boston: Reidel.
Mumford, M. D., Mobley, M. I., Uhlman, C. E., Palmon, P. R., & Doares, L.
M. (1991). Process analytic models of creative capacities, Creativity.
Research J ournal, 4, 91-122.
Mller, R. (2000). Insect visual homing strategies in a robot with analog
processing. Biological Cybernetics, 83, 231-243.
Pollert, L. H., Feldhusen, J. F., Van Mondfrans, A. P., & Treffnger, D. J.
(1969). Role of memory in divergent thinking. Psychological Reports,
25, 151-56.
Root-Bernstein, R. (1988). Discovering. Cambridge, MA: Harvard Univer-
sity Press.
Rubenson, D. L., & Runco, M. A. (1995). The psychoeconomic view of
creative work in groups and organizations. Creativity & Innovation
Management, 4, 232-41.
Runco, M. A. (2004). Creativity. Annual Review of Psychology, 55, 657-
687.
Runco, M. A. (1999c). Time for creativity. Encyclopedia of creativity. San
Diego, CA: Academic, 65963.
Salzberg, C., Antony, A., & Sayama, H. (2004). Complex genetic evolution
of self-replicating loops. Artifcial Life IX: Proceedings of the Ninth
International Conference on the Simulation and Synthesis of Living
Systems (pp. 262-267), MIT Press.
Taylor, C. W. (1988). Various approaches to the defnition of creativity. In
R. J . Sternberg (Ed.), The nature of creativity (pp. 99-124.), Cambridge
University Press.
Thrisson, H. T. (2004). A framework for exploring the evolutionary roots
of creativity. Proceedings of the European Conference on Case-Based
Reasoning 2004 (ECCBR 2004) Workshops (pp. 179-190). Madrid:
Complutense University of Madrid.
van Duijn, M., Keijzer, F., & Franken, D. (2006). Principles of minimal
cognition: Casting cognition as sensorimotor coordination. Adaptive
Behavior,14(2), 157-170.
Weir, A. A. S., & Kacelnik, A. (2006). A new caledonian crow (corvus mon-
eduloides) creatively re-designs tools by bending or unbending aluminium
strips. Animal Cognition. Retrieved September 2006, from http://users.
ox.ac.uk/~kgroup/publications/pdf/WeirKacelnik_06_bend_unbend.
pdf
Wiggins, G. (2006). A preliminary framework for description, analysis and
comparison of creative systems. J ournal of Knowledge Based Systems,
19(7), 449-458
Wolfram, S. (2002). A new kind of science. Champain, IL: Wolfram Me-
dia.
Endnotes
1
See section on Control Network in this chapter & Figure 3 in Thrisson
(2004), fgure not reproduced here.
2
Creature interactions would add another dimension of complexity to
the world.
3
This method of insect associative-memory is often referred to as the
snapshot hypothesis.
4
In prior simulations, we included a mechanism for using an old plan
unmodifed. However, this did not change the overall results in our
experimentation (see Results & Interpretation).
Thrsson
5
Note that we speak of logicality and not of effciency, as plans can
potentially be logical but still ineffcient.
6
This generalization is suffcient for what is discussed here; At higher
levels of complexity, there are quite possibly additional factors to con-
sider.
7
Primitive cognitive systems can fulfll goals without direct knowledge or
understanding that they are doing so, such as the agents in this chapters
simulation.
8
Note that even without the ability to visualize turn of events, a basic
mechanism to produce actions that correlate with the environment serves
the same purpose at a more primitive level.
Sensory Informaton Bottleneck to Central Processng n Adaptve Systems
Chapter.VI
Solving.the.Sensory.
Information.Bottleneck.to.
Central.Processing.in.
Adaptive.Systems
Thomy Nlsson, Unversty of Prnce Edward Island, Canada
Abstract
Information bottlenecks are an inevitable consequence when a complex system
adapts by increasing its information input. Input and output bottlenecks are
due to geometrical limits that arise because the area available for connec-
tions from an external surface always exceeds the area available for connec-
tions to an internal surface. Processing of the additional information faces
an internal bottleneck As more elements increase the size of a processor, its
interface surface increases more slowly than its volume. These bottlenecks
had to be overcome before more complex life forms could evolve. Based on
mapping studies, it is generally agreed that sensory inputs to the brain are
organized as convergent-divergent networks. However, no one has previously
explained how such networks can conserve the location and magnitude of
any particular stimulus. The solution to a convergent-divergent network that
0 Nlsson
overcomes bottleneck problems turns out to be surprisingly simple and yet
restricted.
Introduction
To improve adaptation to the environment by adding sensors, complex sys-
tems need to import the additional information without overcrowding their
processors with more connecting pathways. Any interface between an external
surface and the surfaces of inside structures always faces a potential bottleneck
due to geometry. The external surface can always extend more fnite paths
than can be accommodated by the surface of an internal structure. Biological
systems cannot overcome this bottleneck simply by using smaller connecting
pathways because smaller dendrites and axons conduct nerve impulses more
slowly. More information is useful only if it is also timely.
Adaptation by accessing more information produces bottleneck problems
that do not end once the inputs have been internalized. A slightly different
geometrical constraint now comes into play. The surface area available for
connecting one structure to another grows with the square of each structures
linear dimensions, while the space for elements within a structure increases
with the cube of those dimensions. This can create intraprocessor bottlenecks.
Animals reduce this problem by structuring their processors as sheets. How-
ever, they can not avoid the consequences of additional information requiring
additional processing elements. Larger brains result in longer distances over
which to send information from region to region. The need for information
that is timely requires larger internal pathways, which add to the intraproces-
sor bottleneck problem since space is limited.
Finally, achieving a greater variety of behaviors on the basis of the additional
information may require a greater number of processor outputs to control a
greater number of external effectors. The system now faces a third bottleneck
that is the reverse of the problem encountered by importing more information.
At a more general level, all three types of bottlenecks can be considered as
a single problem of how to connect one set of elements to another set using
fewer connecting paths than the number of elements. Until a solution to this
fundamental bottleneck problem was found, the evolution of more complex
biological systems was stalled.
The evolution of a convergent-divergent organization of animal sensory sys-
tems is generally acknowledged to solve the input bottleneck. This solution
uses spatial multiplexing to convey information from many receptors with
only a fraction as many nerve fbers. Yet there appears to be no explanation
of how such systems can keep track of which receptor is stimulated. This
chapter explains how that is possible.
Neural networks modeled on ordinary computer spreadsheets show how
convergent-divergent organizations of sensory systems can solve the general
bottleneck problem simply by dendritic and axonal branching. They conserve
source location without using tags or feedback. They also convey source
magnitude and are damage tolerant. Trial-and-error modelling suggests that
there is only one basic branching solution for a linear array of sources, and it
is limited to an 8-to-1 convergence. Another solution exists for two-dimen-
sional source arrays, and it is limited to a 16-to-1 convergence.
Hopefully presenting these solutions will lead either to a proof that there is
only one basic method for each type of array and its corresponding limit, or
to other branching solutions. As is, these solutions may contribute to a better
understanding of how information is transmitted in sensory and cognitive
systems and to solving information bottlenecks in other complex systems.
Implications of these convergent-divergent networks for the evolution of
intelligence in biological systems and some possible technical applications
are discussed.
The.Problem
When all else fails, Uncle Remus Brer Rabbit can always escape Brer Fox by
running through the briar patch (Harris, 1955). Brer Fox would be shredded
trying to follow. Real rabbits may not be as smart as Brer Rabbit is portrayed
in these childrens stories, but they share a sensory processing advantage.
Motion pictures of rabbits running through bushes reveal them bobbing and
twisting to dodge the branches. Evidently they have a fne sense of touch that
is rapidly employed. Their survival depends on this ability.
Weddell, Taylor, and Williams (1955) found rabbits a most convenient spe-
cies for studying the peripheral tactile sensory system. The large ears could
be transilluminated for in vivo microscopic examination of the sensory and
Nlsson
neural structures. They found that a square centimeter of skin had some 2000
sensory hair cells. This should give rabbits a fne sense of touch. However,
a moments refection reveals a serious problem. The entire ear would have
about 1/2 million hair cells. How could rabbits get all those nerve fbers from
the ear, not to mention the rest of their body, into a tiny brain and still have
room for some neurons to process the information?
Conceivably, it would be possible using very fne unmyelinated nerve fbers
to transmit the information. However, it is not enough to get the informa-
tion to the brain. It must get there fast to be effective. Table 1 illustrates this
problem. Small unmyelinated dendrites and axons conduct nerve impulses so
slowly that a rabbit running at 20 km/hr would travel nearly a meter before
the touch information reached its brain. Large, fast-conducting axons could
get the information there before the rabbit has travelled a centimeter, but the
diameter of 100,000 such fbers would exceed 30 cm. It is not possible for
rabbits to provide each hair with a suitable pathway of its own to the brain.
This problem is resolved by Weddell et al.s (1955) fnding that there were
only about 6,500 myelinated dendrites entering the ear. Physiological record-
ings from these dendrites revealed that each was connected to some 80 hair
cells. Yet this convergent organization leads to another problem: How can
the brain then tell which hair gets bent? Obviously there has to be some way
or there would be no advantage to having all those hairs.
The means to an answer was revealed as further examination found that each
hair was in fact connected to four dendrites from separate neurons. Since this
FIBER
TYPE
DIAMETER
OF SINGLE
FIBER
DIAMETER
OF 100,000
FIBERS
IMPULSE
SPEED
DELAY
DISTANCE
TRAVELLED
@ 20 km/hr
(:m) (cm) (m/sec) (sec) (cm)
SMALL: 0.5 1 0.5 0.15 83
LARGE: 20 >30 90 0.002 <1
Table1.Diameterofanervebundlerequiredtocarryanervefberforeach
ofthe100,000hairsonarabbitsear,andhowfararabbitwouldtravelat
20km/hrduringthedelaybetweenwhenahairistouchedandtheinforma-
tionreachesitsbrainforsmallandlargenervefbers.
divergence of information from each receptor reduces the overall convergence
from 80-to-1 to 20-to-1, the divergence must be necessary. A clue to the role
of divergence arises from their observation that the dendrites connected to any
given hair cell followed diverse routes to exit the ear. Weddell et al. suggest
that divergent paths avoid cross-talk between fbers responding at the same
time. Brown et al. (2000) fnd that cats have a similar convergent-divergent
organization of their peripheral tactile system. Continuation of this conver-
gent-divergent organization into the somatosensory cortex is related to spatial
acuity of cats (Brown, Koerber, & Millecchia, 2004) and to the fne grain
of somatosensory receptive felds in monkeys (Rausell, Bickford, Manger,
Woods, & J ones, 1998). Thus, the brain must somehow use the responses
in many divergent fbers to determine which hair gets bent despite the high
degree of peripheral convergence.
The problem is not peculiar to the tactile system. The human eye has over
100 million photoreceptors, but the optic nerve has only some 1.2 million
axons (Polyak, 1957). Even in the fovea where there is a one-to-one ratio of
receptors to ganglion cells, there is both convergence of several receptors
connecting to a single ganglion cell and divergence of each receptor con-
necting to several ganglion cells (Dacey, 2004). Nevertheless, visual acuity
can equal and even be fner than the width of a single receptor (e.g., Blake
& Sekuler, 2006).
Background
Convergent-Divergent.Networks

To solve the information bottleneck, a convergent-divergent organization of
neural branching is generally acknowledged as a fundamental property of
most sensory systems (e.g., Guyton, 1991; Rosenzweig, Breedlove, & Wat-
son, 2005; Uttal, 1973). In the periphery, convergence predominates with
the number of receptors greatly exceeding the number of afferent pathways.
Once sensory information passes the thalamus, a predominantly divergent
organization of connections from these pathways to many cortical neurons
restores receptor specifcity. Figure 1 illustrates the general concept of a
convergent-divergent system. Upon reaching the cortex, information con-
veyed by a convergent-divergent sensory system somehow indicates which
Nlsson
receptor was the source. It is certainly not obvious how convergent-divergent
branching reveals the source, and the acknowledgments do not explain how
this can happen.
How can receptor location be maintained through the bottleneck of a few
connecting fbers? Any general answer pertaining to biological systems must
recognize that all nerve impulses are the same regardless of the initiating
source. While temporal patterns of nerve impulse trains could encode receptor
specifcity, this seems an unlikely solution to the bottleneck problem since
temporal patterns are needed to carry qualitative information such as magnitude
of receptor response. It is also possible that feedback loops could be used to
decode which receptor is active as suggested by Goldberg, Cautwenberghs,
and Andreou (2001). However, reliance on feedback would delay using the
information and require more connections to the periphery.
This chapter explains how convergent-divergent encoding and decoding
can transmit information from many receptors across a fractional number
of connecting paths simply by using certain neural branching patterns and
Figure 1. The general concept of a convergent-divergent information trans-
mission system. Many INFORMATION INPUTS are provided to some type of CON-
VERGENT ENCODER, which processes the information and sends it along a few
CONNECTING PATHWAYS. A distant DIVERGENT DECODER spreads the information
from the PATHWAYS to as many OUTPUTS as there are sources. The information
from any particular INPUT goes to an OUTPUT corresponding to that source.
(copyright by author, printed with permission)
connecting weights. Such nets passively achieve a form of spatial multiplex-
ing without either adding information to identify the sources or waiting for
feedback confrmation.
Conserving.Input.Location.
Compacting information from many receptors simply by connecting them
to a single afferent neuron would make it impossible to tell which receptor
was being stimulated. Convergence must be accompanied by divergence to
conserve receptor identity. Both types of branching are evident in the af-
ferent connections of the retina and lateral geniculate nucleus of the visual
(Reid & Usrey, 2004) and tactile (Brown et al., 2000) sensory systems. The
key is a neural branching pattern that has a different number of connections
or branch weights between each receptor and two or more afferent neurons.
Physiologically, different weights can be achieved by the number of dendritic
branches and by the number of pre- and post-synaptic structures within a
single synapse. Weighted branching can produce different response levels in
two or more afferent fbers depending on which receptor is stimulated.
Now consider the divergence of connections from the afferent nerve fbers
to a larger set of cortical neurons that correspond to individual receptors.
That divergence must be accompanied by some convergence so that each
afferent fber connects to more than one cortical neuron. Otherwise, all
cortical neurons connected to the same afferent fber would have the same
outputnothing would be accomplished by the divergence. Both convergent
and divergent branching is evident in connections from the lateral geniculate
to the visual and somatosensory cortex (Alonso & Swadlow, 2005). Mirroring
the branching pattern for convergent encoding, each cortical neuron receives
a different number of branches from each afferent fber. This produces a dif-
ferent response strength in a cortical neuron depending on which afferent
fbers are active.
In response to stimulation of a single receptor input, the output of such con-
vergent-divergent systems is more complex than what is produced when each
receptor has its own pathway. Divergence results in many cortical neurons
responding to stimulation of a single receptor. Convergence results in a dif-
ferent cortical response strength depending on which receptor is activated.
With the correct branching organization, the cortical neuron with the maxi-
mum response will be the one whose location corresponds to the location of
the stimulated receptor.
Nlsson
Conserving.Input.Strength
The ability of convergent-divergent branching networks to also transmit the
magnitude of an input is essential to convey what has happened as well as
where something has happened (e.g., Sheth & Shimojo, 2002). Neurons must
convey stimulus strength by temporal patterns of impulsestypically fring
faster to stronger stimuli. Since branching solutions just use multiplication
and summation, doubling the stimulus strength not only doubles the output
response but also doubles the difference between the maximum response
and the next strongest response. Thereby convergent-divergent systems can
conserve both receptor identity and stimulus magnitude.
Solution
A solution to the bottleneck problem occurred spontaneously in response to
a students question about the general convergent-divergent organization of
sensory pathways. Intuitively, I sketched on the blackboard a simple linear
array of inputs connected by a 2 and 1 branching pattern to the same number
outputs with one pathway for every two receptors. Stimulating one receptor
and counting the number of active intersections leading to each output pro-
duced a maximum at a location that corresponded to the input location. That
went well, so I added sketch an example to my notes. Thinking that any
symmetrical pattern of convergent and divergent branching would work, it
seemed unnecessary to be more specifc. A few years later, I sketched a pat-
tern, but was embarrassed to fnd it did not work. Trying to get a branching
pattern that would work, revealed that symmetry was not suffcient. Indeed
many tedious sketches produced only one pattern that did work.
Some years after that I programmed this pattern in MicroSaint (1985). It
worked but ran too slowly to be practical either for demonstration or fur-
ther study. When classroom computer projectors became available, I made
probability demonstrations for my statistics course using QuattroPro 4
(1987) spreadsheets. Could a convergent-divergent network be modeled on
something as simple as a spreadsheet? Yes, and this enabled testing many
confgurations. Remarkably, there did seem to be only one basic pattern for
successful convergent-divergent networks. Had something this basic not
been discovered previously? Presented at a meeting of the Association for
Mathematical Psychology (Nilsson, 2002), the model raised interested com-
ments which suggested the solution was indeed novel.
Preparing a fgure to explain the linear array model for this chapter, led to the
idea to try making a convergent-divergent model that would handle inputs
arrayed over a 2-dimensional surface. This would provide a more realistic
solution to the sensory bottleneck problem encountered in the tactile and
visual systems. The solution to a convergent-divergent system that operates
on a 2-dimension array of inputs requires a 3-dimensional neural network.
It will be easier to understand how this model works by frst considering the
principles of its operation using a model that operates on a linear array of
inputs and outputs.
Linear.Array.Model
Basic Operation
To see how a convergent-divergent network can work just by branching
weights, consider a linear array of n inputs with n/2 connections to a linear
array of n outputs. Figure 2 shows how 8 input cells B, C, ..., I of such a
network can send information to 8 output cells, B, C, ... I, using 4 con-
necting pathways, BC-BC, DE-DE, FG-FG, and HI-HI. There are twice
as many input cells as convergent cells BC, DE, ... HI. Each input cell sends
two branches to the nearest convergent cell and a single branch to the next
nearest convergent cell. Each output cell receives two branches from its
closest divergent cell and a single branch from the next closest divergent
cell. That is all that is required.
In Figure 2, the different number of branches between cells is literally indi-
cated by single or double paths. Each branch independently conveys the full
value of the response strength in its originating cell, which lies to the left in
the diagram. (Subsequently it will be necessary to refer to the strength of such
connections as weights to simplify explanation.) The strength of the response
in any cell is the sum of what is received from the incoming branches. Thus
convergent cells DE and FG have the values 2 and 1 respectively. These
values are directly transmitted over the connecting pathways to divergent
cells DE and FG. Because output cell E receives two inputs of 2 from
Nlsson
Figure2.Schematicforalineararrayofinputsandoutputswithatwo-to-
one and one-to-two convergent-divergent system. It sends information along
half as many CONNECTING PATHWAYS as there are INPUTS and OUTPUTS. Entering
a value of 1 into INPUTcellEproducesthevalues0,2,4,5,4,2,1,and0
at OUTPUTS B, C, D, E, F, G, H and I respectively. The OUTPUT with the
largest response corresponds to the INPUT that received the information. The
COMPARISON NETWORK only passes on the OUTPUT with the maximum value, which
in this example is the 5 from E. Each path from a cell represents a neural
branch that conveys the value in that cell. Note that some cells such as E and
DEareconnectedbydoublelines-thusDEhasthevalueof2sinceithas
two branches from E each of which conveys Es value of 1. Note: lines that
cross remain independent. (copyright by author, printed with permission)
DE plus an input of 1 from FG it has the strongest output, 5. The location
of this maximum output at E corresponds to the location of the input cell E
which received the stimulus.
Implementing the previous on a spreadsheet requires viewing matters from a
different perspective. It is not possible for two or more cells in a spreadsheet to
send their contents to a third cell and expect to fnd the sum of their contents
in that cell. Rather the third cell must be programmed to add the contents of
the two sending cells. This results in a subtle but critical change in creat-
ing the convergent part of the network. From the perspective of information
fow, branching from the input cells is always asymmetrical: two branches to
the nearest convergent cell and one branch to the next-nearest. Yet from the
perspective of the convergent cell where the mathematical operation resides,
the incoming branching is always symmetrical: two branches from each of
the two closest inputs and one branch from each of the two next-furthest in-
puts. The opposite occurs in the divergent section. Here the divergent cells
send a symmetrical pattern of branches to the output cells. However, the
formulas in the output cells require an asymmetrical summation from the
nearest and next-nearest divergent cell.
Winner-Take-All Economics

Once a convergent-divergent system has achieved a distribution of output
responses with a maximum at a certain cell that corresponds to the input, the
system has achieved its task of overcoming the information bottleneck. From
the topography of the maxima in such distributions, various simple and more
complex receptive felds can be assembled to achieve the sensory processing
for lines, edges, spatial frequency, etc. A general output of such distributions
could suffce in principle to having transmitted the location and magnitude
of the stimulus. However, an output that consisted only of the maxima would
simplify the subsequent processing to produce cortical receptive felds and
refex arcs. That can be achieved by connecting the convergent-divergent net-
work outputs to a comparison network that emits only the maximum output
at its location. Such a network is readily obtained by lateral inhibition that
produces a winner-take-all effect over a range that may be related to atten-
tion processes (Maass, 2000). This is illustrated by the right hand comparison
network in Figure 2. Each cell in that layer produces an output if its value
exceeds the value of its neighboring cells. (The indicated calculation, IF
0 Nlsson
(B(A...C), B, 0) represents the logic using minimal space. However, in
QuattroPro the operation must be replaced by =@MAX.)
Response Magnitude and Tolerance
In the spreadsheet model, input and output magnitude are represented as
numbers. Figure 2 shows that an input value of 1 at E produces a maximum
output of 5 at E. Simple arithmetic along the branches shows that an input
of 2 produces a maximum output of 10; an input of 3 produces an
output 15; etc. Not only is the maximum output proportional to the input
by a factor of 5, the difference between the maximum output and the next
largest output equals the input. This shows that the convergent-divergent
branching network can directly transmit an inputs magnitude as well as its
location.
Convergent-divergent networks are damage tolerant. In the example of Figure
2, if either of the two active connecting pathways, DE-DE or FG-FG, are
disabled, a maximum still appears at outputs that are close to the output that
corresponds to the location of the input. With an input 1 at E, disabling
DE-DE produces two output maxima of 2 at F and G; disabling FG-FG
produces maxima of 4 at D and E. Networks with more layers of con-
vergent and divergent branching (see below) are more tolerant of disabled
connecting paths since they activate more paths for any given input.

More Convergence
Instead of connecting the convergent cells BC, DE ..., HI to divergent cells
BC, ..., HI as shown in Figure 2, they could function as input cells to a
second layer of convergent cells which one might label BCDE, FGHI, etc.
This together with a corresponding second layer of divergent cells, BCDE,
FGHI, ..., produces a convergent-divergent network with one connecting
pathway for every four inputs and outputs. Likewise adding a third layer of
convergent and divergent cells produces an 8-to-1 convergence ratio. See
Figure 3. The effect of such cascaded layers can also be merged into single
convergent and divergent layers with an equivalent, broader pattern of
weighted branches. Because the resulting schematic needs to be large and is
diffcult to interpret even with colored paths to indicate the various number
of branches, a fgure to illustrate this possibility was not feasible.
Can any number, x, of convergent and divergent layers be cascaded to achieve
2
x
convergence ratios? No. Adding a fourth or more set of layers does not
work. The output cell with the maximum response ceases to correspond with
the input cell that is stimulated. The error patterns are predictable in terms of
where the input stimulus occurs, but using that information still leaves some
ambiguity and begs the purpose of the network.
Figure3.Threelayersofconvergentandofdivergentcellscanbecascaded
toproducean8-to-1convergenceandstillconservestimuluslocationand
magnitude. (copyright by author, printed with permission)
Nlsson
Other Patterns and Limits?
Other than combinations of the binary, 2 and 1 branching pattern described
above. I could not fnd another basic convergent-divergent pattern that cor-
rectly transmitted the location of an input. For example, there seem to be
no solutions that produce 3-to-1, 5-to-1 or other even but not power-of-2
convergence ratios such as 6-to-1. Nor could I fnd any variation of 2 and 1
branching that overcame the limit of an 8-to-1 convergence. Perhaps someone
will either prove me wrong or fnd a general explanation for why only the
previous pattern works and its inherent limit.
Acuity
Explanation so far has involved only a single input into the convergent-di-
vergent network. When two or more inputs occur simultaneously, the results
depend on the distance between the inputs and the convergence ratio. The
2/1 convergence network shown in Figure 2 can accurately locate two inputs
of equal strength as long as they are separated by four cells. Smaller gaps
produce one or two maxima at outputs between those that correspond to the
inputs. The triple convergence-divergence network shown in Figure 3 with 8-
to-1 convergence does not accurately locate two inputs until they are 16 cells
apart. Clearly networks that achieve accurate convergence and divergence
simply by branching, do so at the expense of ability to resolve simultaneous
inputs. (How this liability may be bypassed is discussed later.)
2-Dimensional.Arrays
Basic Model
To model a convergent-divergent neural network that would operate on an
array of inputs spread over a surface, I initially copied a spreadsheet version
of Figure 2 to every other page in a 3-dimensional spreadsheet program,
QuattroPro12. The convergent and divergent calculations were moved to
the in-between pages and connections to input and output cells extended to
both adjacent sheets. This proved awkward. So instead I placed the inputs as
an input array on one page, the convergent calculations on the next page,
the connecting pathways on a 3rd page, the divergent calculations on a
4th page, and the outputs along with comparison network calculations as an
output array on a 5th page. (It can all be done on just two pages, one for
convergent calculations and one for the divergent, but having additional pages
for inputs, pathways and outputs was easier to follow.) Figure 4 illustrates
this organization for a convergent-divergent network operating on a sheet of
inputs with a single level of convergent and divergent calculations.
It was immediately evident that the 2 and 1 branching used for linear arrays
was no longer appropriate. Each input cell now has four adjacent convergent
calculation cells to which it must send branches. To make the closest and
next-closest branching weights inversely proportional to their distance required
three weights that could no longer be simple whole numbers. Pythagoras
theorem was used to calculate the vector lengths or Euclidian distance, and
the reciprocal of that distance became the weight for that connection.
(There is no distance between sheets. Decreasing the branch weights in pro-
portion to distance-squared does not work. Therefore, the spread of information
in these neural networks is not analogous to a radiant effect. This indicates
that convergent-divergent networks involve a selective branching structure
as opposed to a simple spreading of the branches, which would produce a
radiant-like reduction in terms of branch density as distance increased.)
As shown for the convergent cells on Page 2 of Figure 4, the response
from each input cell is multiplied by one of these branching weights and
summed with the results from the other input cells connected to a given
convergent cell. It is easier to visualize the weights from the perspective of
the convergent cell because they are spatially symmetrical and the same for
all convergent cells. Figure 5 shows the weights of the branches from the
sixteen input cells that send branches to each convergent cell. Note how a
third weight of 0.24 is required for the next-closest set of input cells C4, I4,
C11, and I11 located at the corners of the twelve cells surrounding four cells
closest to the convergent cell at F7.
On Page3, the convergent cell locations become the connecting pathways
between the peripheral input array cells and the central processor that lies
beyond the output array cells that form the limits of the convergent-diver-
gent transmission model. The weights of the branches from these pathways
to the output array cells on Page 5 look the same as the weights shown in
Figure 5. The difference is that the centrally located pathway at F7 would be
connected to output array cells at C4, E4, G4, ..., I11. However, this sym-
metric branching pattern can not be implemented in the spreadsheet model
since spreadsheet cells can not send their contents to other cells.
Nlsson
Figure4.Howaneuralnetworkforatwo-dimensionalarrayofINPUTS and
OUTPUTSwithsingledegreeof4to1convergenceand1to4divergencewas
modeledonfvespreadsheetpages.ExceptfortheINPUT cells on Page 1, each
active cell is connected to cells on the previous page by neural branches that
havevariousweightssuchastheindicatedvalues0.71,0.34,etc.usedinthe
CONVERGENT and DIVERGENT CALCULATIONSonPages2and4.Theseweightsreduce
the strength of what is received from the branchs source cell. Stimulating the
INPUTcellatlocationG8onPage1producesamaximumresponseatlocation
G8onPage4.IneachindicatedcellonPage5anIFfunctionoutputsthe
responsevalueofthecorrespondingcellonpage4ifthatvalueislargerthan
any of its neighbors. Page 1 and Page 5 are only connected by the PATHWAYS
shownonPage3.(copyrightbyauthor,printedwithpermission)
The divergent cells on Page4 of the network in Figure 4 operate by add-
ing the contents from their two closest and two second closest connecting
pathways. The branches from each of these pathways require three different
weights so that the contribution from each pathway is inversely proportional
to its distance from the output cell. See Figure 6. With respect to output G8
in the upper illustration, the branch from its closest pathway, F7, has a weight
of 0.71; the two next closest pathways, F11 and J 7, are weighted 0.32; and
the furthest, J 11, is weighted 0.24. Note that the weight of branches from
each of these pathways varies depending on which output is receiving their
branches. This is shown in the lower illustration for output I10.
Figure 5. The weights used by CONVERGENT CALCULATION cells on Page 2 of
Figure 4 for branches from Page 1 INPUTS to Page 3 CONNECTING PATHWAYS
(shown shaded). The indicated weights are those used by CONVERGENTcellF7
for branches from INPUTSC4,C6,...,I10toPATHWAYF7.Theweightforeach
branch is the reciprocal of that INPUTS distance to the PATHWAY. (Consider
Page2tohavezerothickness.)EveryCONVERGENT cell uses the same weights
for its INPUTS. (copyright by author, printed with permission)
Nlsson
Figure 6. Weights used by the DIVERGENT CALCULATIONcellsonPage4ofFigure
4forneuralbranchesfromtheCONNECTING PATHS(shownshaded)onPage3to
the OUTPUTS on Page 5. There are four sets of asymmetrical weights depend-
ing on an OUTPUT cells position with respect to its nearest PATHWAY. The upper
diagram shows the weights for OUTPUT E6 of the branches from the adjacent
PATHWAYF7andthethreeotherPATHWAYS(B3,F3,andB7)thatsurroundF7.
The lower diagram shows the weights for OUTPUTG8.(copyrightbyauthor,
printed with permission)
Similar to the linear array model, the 2-dimensional output array of a conver-
gent-divergent network is a distribution of response values across the array.
Only the maximum value of that distribution corresponds to the location of
an input. Figure 7 shows the distribution of values in the output array prior
to any comparative processing when a 1 has been entered at input G8.
The maximum value in this array correspondingly occurs at output G8. The
values in other output cells decrease in proportion to their distance from
this location.
Figure7.Thedistributionofoutputvaluesacrossaportionofthe2-dimen-
sional OUTPUT ARRAYonPage5ofFigure4priortothewinner-take-allcalcu-
lations in response to an input value of 1 entered at INPUTG8onPage1.
Notethatthemaximumvalue,0.76,occursatOUTPUTG8,whichcorresponds
to the location where the 1 was entered in the INPUT ARRAY. Only CONNECT-
ING PATHWAYSF7,J7,F11,andJ11contributetotheseoutputsaccordingto
their distance from each OUTPUT cell as shown in the upper part of Figure 6.
(copyright by author, printed with permission)
Nlsson
As illustrated in Figure 4, a 3-dimensional neural net can transmit information
from a sheet of input cells along one-quarter as many connecting pathways
as the number of inputs and accurately represent in its output array which
input was stimulated and the magnitude of that input. The cost of achiev-
ing this compressed transmission is that additional processing is needed to
identify which output has the information that specifcally corresponds to
the input. Subsequent winner-take-all calculations such as the IF(G8 (C4...
K12), G8, 0) shown at output G8 on Page 5 of Figure 4 are needed at each
output to isolate the location and value of the output cell having the largest
value - in this example 0.76 at location G8.

Some Further Properties
As with the linear array model, greater convergence is possible by adding
more layers of convergent-divergent branching before and after the connecting
paths. On Page2 of Figure 4, instead of having the branches from the input
cells go directly to the connecting pathways, their destination can be cells
with a second set of similar convergent calculation cells which then branch
to the pathways. This double convergence uses only one-quarter as many
pathways as the network in Figure 4. This results in a 16-to-1 convergence.
A comparable doubling of the divergent cells in Page4 still conserves input
location and magnitude.
A third layer of convergent and divergent cells on Page2 and Page4 would
produce 64-to-1 convergence. However, errors now appear in localizing the
input. Changing the branching weights changes where the errors occur with
respect to the center and periphery of a 64 cell unit matrix of output cells.
As weights are changed, errors appear in the center of this matrix before they
disappear in the periphery. This suggests that there is no solution. Sixteen-
to-one convergence appears to be the practical limit for a 3-dimensional
branching network operating on a sheet of inputs and outputs.
As could be expected, the spatial acuity of these 2-dimensional networks is
closely related to the degree of convergence. The single degree 4-to-1 net-
work shown in Figure 4 will locate two simultaneous inputs down to a gap
of 4 input cells. A second set of convergent and divergent cells will locate
two inputs down to a gap of 14 cells.

Emerging.Insights
Biological.Implications
The 16-to-1 convergence obtainable simply by neural branching enables a
substantial reduction of neural pathways needed to transmit information from
many sources while maintaining the origin and magnitude of the information.
Sensory systems such as the rabbit somatic system studied by Weddell et al.
(1955) have a convergent ratio of 20-to-1 which is similar to the 16-to-1 ratio
obtainable with the above branching model. Such savings in overall pathway
size would also be applicable to major intracortical tracts which also must
present bottlenecks to the fow of information between cortical areas. The
neural branching evident at each end of such tracts seems capable of sup-
porting a convergent-divergent organization (Bullier, 2004). Though these
pathways also seem to involve feedback, that may serve other purposes such
as refning and modifying the transmitted information rather than ensuring
its accuracy.
Modelling these convergent-divergent systems may appear a daunting task
and lead one to wonder how they could ever be achieved precisely enough
biologically. In practice, once the basic branching pattern is defned, the rest
is simply a matter of large scale copying easily done on spreadsheets. A gene,
Pax6, has been found to not only initiate the development of eye structures
(Kozmik, 2005), but also produces a topographic network of central con-
nections (Vanderhaeghen & Polleux, 2004). The operation of this gene has
been traced back to Coelenterates (jellyfsh)the simplest animals having
a visual system (Nilsson, 2004). This suggests that all visual systems share
a common origin dating back to the Precambrian era. While a parallel or-
ganization of individual receptor nerve fbers seems to have been adequate
for simple animals, at some evolutionary point visual and tactile systems
with multitudes of receptors and larger, more complex bodies must have
encountered the bottleneck problem of how to convey all that information
into a central nervous system. Then it is likely that a convergent-divergent
variation of earlier parallel pathway systems was found adaptive.
However, more complex animals that adapted a convergent-divergent system
faced a problem not present in parallel systems. In response to a single input,
convergent-divergent networks produce a distribution of outputs that vary
in magnitude. In the previous models, a layer of comparison calculations
0 Nlsson
was added at the output layer to emit only the maximal output, which cor-
responds to the input. Neural network implementation of such calculations
use lateral inhibition to achieve a winner-take-all effect (Maass, 2000). A
single-input-to-single-output system would facilitate obtaining the refex-arc
learning of classical conditioning models such as those proposed by Sher-
rington, Pavlov, and Hebb (Hebb, 1949) that seem realizable with long-term
potentiation effects (Bliss & Collingridge, 1993). Therefore, convergent-di-
vergent sensory transmission may have provided the impetus for developing
information comparison networks.
Convergent-divergent comparison networks could be capable of much more
than facilitating refex arcs. The combination of multiple interconnected in-
formation pathways with inhibitory comparison processing is similar to the
networks proposed by Pribram (1991) to achieve holographic information
processing which shares many properties associated with consciousness.
Adding memory to such circuits can produce recognition systems (Deutsch,
1967; Eich, 1985; McLeod, Plunkett & Rolls, 1998). Might the convergent-
divergent organization of the visual system have been the basis for evolving
more complex information processing? Vertebrates are unique in having a
visual nervous system that is part of their central nervous system. In mam-
mals, the visual system develops in the neural tube prior to the forebrain.
Technological.Possibilities

Branching solutions to spatial multiplexing seem applicable to electronic,
optical, and even hydraulic information systems. Wire, optical fbers, beams,
or pipes are information transmission media that are readily branched, and
the detection and measurement of maxima in these media does not seem
diffcult. Multiplexing by means of branching directly in these media may
prove advantageous to operations that must continuously adapt to changing
conditions. Here are some specifc application possibilities:
1. Like animal sensory systems, microcomputing circuits encounter bottle-
necks whenever information has to be passed from one component to
the next. Temporal multiplexing is presently used to overcome such
bottlenecks. An ability to also incorporate spatial multiplexing could
reduce the bottleneck problem by an additional factor of 16. This would
seem important to enabling computing components to handle more
information.
2. Temporal multiplexing incurs a penalty of reducing operating speed by
the multiplexing factor. For example, if an information transmission
path has to be switched 4X to send the entire message, the fnal rate at
which the complete information can be used is reduced by a factor of
4. Spatial multiplexing operates immediately. It does not need to take
time out to break down a message into chunks for sequential transmis-
sion followed by reassembly. Therefore spatial multiplexing would
be advantageous over temporal multiplexing to increase the speed of
information processing.
3. Spatial multiplexing can handle analog information directly. The relative
strength of the output is proportional to that of the input. While tempo-
ral multiplexing can in principle also handle analog signals, switching
introduces transients, which distort analog signals.
4. The analog ability of spatial multiplexing would, in principle, enable it
to work with any kind of information media: optical, hydraulic, acous-
tic, pneumatic. This suggests that spatial multiplexing could be directly
applied in non-electronic control systemsfor example to transmit
hydraulic operations to a multitude of actuators along a few hydraulic
lines.
Spatial.and.Temporal.Tradeoffs

The major drawback to saving transmission paths by branching is the dif-
fculty in handling more than one input simultaneously without unacceptable
loss of spatial acuity. Since this problem has been resolved in the complex
sensory systems of vertebrates, this may not be as severe a problem as it frst
appears. Notwithstanding inescapable neural conduction latency, the tempo-
rally discrete nature of neural impulses is smoothed by the combination of
asynchrony of individual neural responses and variations in neural conduc-
tion speed. Multiplexing achieved by branching is continuous and does not
require processing time. This can be a critical advantage when coherency is
required for information processing (Cotterill, 1988). Continuous transmission
systems can be considered to have infnite temporal resolution. The common
wisecrack, God invented time to prevent everything from happening at
once. has signifcance here. Strictly speaking, no two things probably ever
happen at exactly the same time. Thus a system that can transmit information
continuously need not lack spatial acuity after all.
Nlsson
Conclusion
The bottleneck problem of how to transmit information from many sources
without an overwhelming number of connecting pathways is common to most
complex systems that seek more information to adapt to their environment.
The problem is compounded in biological neural systems where transmission
delays interact with pathway diameter and system size (Ringo, Doty, Demeter,
& Simard, 1994). Unsolved, this bottleneck places severe restrictions on the
viable size and complexity of adaptive systems.
Two biologically inspired neural network models for a convergent-divergent
organization of an information transmission system have been presented. They
enable information from many inputs to be transmitted along a fractional
number of connecting paths while conserving which input was the source of
the information. They solve the bottleneck problem using certain branching
patterns or weights of the connections to and from the paths that transmit
information from inputs to outputs. Their branching pattern makes them
more tolerant of component failure than parallel transmission systems. The
continuous operation of these models enables them to maintain the magnitude
of the input and minimize transmission delays.
The inherent lack of complexity in how the elements of these models operate
shows that spatial multiplexing can be achieved by neural networks that are
as simple as models by McLeod et al. (1998) that are capable of learning and
recognition. Adding these convergent-divergent transmission principles to the
front-end of such cognitive models will enable future designers of complex
adaptive systems to more realistically maximize their information handling
abilities. These principles can also solve the need for biological modelling
of how information from many sources can be economically transmitted
from one region to another within a brain (Karbowski, 2003) and also from
a brain to operate a large number of effectors such as individual muscle cells
(Prescott, 2007).
In summary, solving the sensory information bottleneck has revealed the fol-
lowing nine general principles and a tenth speculative principle that enable
complex adaptive neural systems to use more information:
1. Obtaining more information from the environment requires connecting
more sensors to the central processor. Processing the additional infor-
mation requires a larger processor. Using the information may involve
sending more information to control behavior.
2. To be useful, the information must arrive promptly. This becomes more
diffcult as neural systems get larger and information must travel fur-
ther.
3. Biological neural systems are constrained by a tradeoff between con-
ductor size and information conduction speed. Larger neurons conduct
more rapidly, can support more branches, but take up more space than
smaller neurons.
4. Central processing of information is advantageous over distributed
processing because it facilitates rapid integration and comparison of
information from diverse sources, including memory, by reducing the
distances information must travel.
5. At some point in increasing the number of information pathways and
consequent growth of system size, the number of information path-
ways that can enter a central processor faces a bottleneck due to space
limitation in how many connections are possible at the surface of the
processor. Faster biological pathways exacerbate this space limitation
by being larger.
6. Similar bottlenecks exists within a processor when large amounts of
information need to be rapidly conveyed from one region to another
and when large amounts leave the processor to control behavior.
7. These bottlenecks can be solved by certain convergent-divergent branch-
ing organizations of the neural pathways. These organizations enable
information from many sources to be conveyed by fewer connections
while conserving the origin and magnitude of the transmitted informa-
tion. (Biologically, the required additional branching is readily obtained
as larger neurons are used to ensure promptness.)
8. Unlike parallel transmission systems where information from each source
is contained within a single pathway, convergent-divergent information
networks distribute information from each source across many pathways
that also carry information from other sources.
9. Economical use of this distributed information may require the addi-
tion of comparison networks that identify the location and value of the
distributions maximum which identifes the original input in terms of
a particular neural element. This facilitates combining the information
from several inputs to form specialized elements such as receptive
felds. It also facilitates channelling an input into refex arcs and other
behaviors.
Nlsson
10. (speculative). Modest modifcation of the blueprint for building
comparative convergent-divergent information transmission networks
may be combined with memory to produce recognition, learning, and
consciousness.
References
Alonso, J. M. & Swadlow, H. A. (2005). The thalamocortical specifcity
and the synthesis of sensory cortical receptive felds. Journal of Neu-
rophysiology,94,26-32.
Blake, R., & Sekuler, R. (2006). Perception. Toronto: McGraw-Hill.
Bliss, T. V. P., & Collingridge, G. L. (1993). A synaptic model of memory:
Long-term potentiation in the hippocampus. Nature, 361, 31-39)
Brown, P. G., Harton, P., Millecchia, R., Lawson, J ., Kunjara-Na-Ayudhya,
T., Stephens, S., Miller, M. A., Hicks, L., & Culberson, J . (2000). Spa-
tial convergence and divergence between cutaneous afferent axons and
dorsal horn cells are not constant. Journal of Comparative Neurology,
420, 277-290.
Brown, P. G., Koerber, H. R., & Millecchia, R. (2004). From innervation
density to tactile acuity, I. Spatial acuity. Brain Research, 1011, 14-
32.
Bullier, J. (2004). Communications between cortical areas of the visual system.
In L. M. Chalupa & J . S. Werner (Eds.), The visual neurosciences. (pp.
522-540). Cambridge, MA: MIT Press.
Cotterill, R. M. J . (1988). A possible role for coherence in neural networks.
In R. M. J . Cotterill (Ed.), Computer simulation in brain science (pp.
164-188). Cambridge: Cambridge University Press.
Dacey, D. (2004) Origins of perception: Retinal ganglion cell diversity and
the creation of parallel visual pathways. In M. S. Gazzaniga (Ed.),
The cognitive neurosciences III. (pp. 281-301). Cambridge, MA: MIT
Press.
Deutsch, S. (1967). Models of the nervous system. New York: J ohn Wiley
& Sons.
Eich, J. M. (1985) Levels of processing, encoding, specifcity, elaboration,
and CHARM. PsychologicalReview,92(1).
Goldberg, D. H., Cautwenberghs, G., & Andreou, A. G. (2001). Probabilistic
synaptic weighting in a reconfgurable network of VLSI integrate-and-
fre neurons. Neural Networks, 14, 781-793.
Guyton, A. C. (1991). Textbook of medical physiology. Philadelphia: W. B.
Saunders.
Harris, J . C. (1955). The complete tales of Uncle Remus. Boston: Houghton
Miffin.
Hebb, D. O. (1949). Organization of behavior. New York: Wiley.
Karbowski, J . (2003). How does connectivity between cortical areas depend
on brain size? Journal of Computational Neuroscience, 15, 347-356.
Kozmik, Z. (2005). Pax genes in eye development and evolution. Current
Opinion in Genetics and Development, 15(4), 430-438.
Levine, M. W. (2000). Fundamentals of sensation and perception. Oxford:
Oxford University Press.
Maass, W. (2000). On the computational power of winner-take-all. Neural
Computation, 12, 2519-2535.
McLeod, P., Plunkett, K., & Rolls, E. T. (1998). Introduction to connectionist
modelling of cognitive precesses. Oxford: Oxford University Press.
MicroSaint [Computer program] (1985). Boulder, Colorado: Micro Analysis
and Design.
Nilsson, D. E. (2004). Eye evolution: A question of genetic promiscuity.
CurrentOpinioninNeurobiology,14, 407-414.
Nilsson, T. H. (2002). How acuity is conserved in convergent-divergent
networks. Paper presented at the meeting of the Association for Math-
ematical Psychology, Miami University, Oxford, Ohio.
Polyak, S. (1957). The vertebrate visual system. Chicago: University of
Chicago Press.
Prescott, T. J . (2007). Forced moves or good tricks in design space? Landmarks
in the evolution of neural mechanisms for action selection. Adaptive
Behavior, 15, 9-31.
Pribram, K. H. (1991). Brain and perception. Hillsdale, NJ : Lawrence Erl-
baum Associates.
Nlsson
QuattroPro 4 [Computer program] (1987) Scotts Valley, California: Borland
International.
QuattroPro 12 [computer program] (2006) Ottawa: Corel Corporation.
Rausell, E., Bickford, L., Manger, P. R., Woods, T. M., & J ones, E. G.
(1998). Extensive divergence and convergence in the thalamocortical
projection to monkey somatosensory cortex. Journal of Neuroscience,
18, 4216-4232.
Reid, R. C., & Usrey, W. M. (2004). Functional connectivity in the pathway
from retina to striate cortex. In L. M. Chalupa & J . S. Werner (Eds.), The
visual neurosciences. (pp. 673-679). Cambridge, MA: MIT Press.
Ringo, J . L., Doty, R. W., Demeter, S., & Simard, P. Y. (1994). Time is of
the essence: A conjecture that hemispheric specialization arises from
interhemispheric conduction delay. Cerebral Cortex, 4, 331-343.
Rosenzweig, M. R., Breedlove, S. M., & Watson, N. V. (2005). Biological
psychology. Sunderland, MA: Sinauer Associates.
Sheth, B. R., & Shimojo, S. (2002). Signal strength determines the nature
of the relationship between perception and working memory. Journal
of Cognitive Neuroscience, 15, 173-184.
Uttal, W. R. (1973). The neurobiology of sensory coding. New York: Harper
& Row.
Vanderhaeghen, P., & Polleux, F. (2004). Developmental mechanisms pat-
terning corticothalamic projections: Intrinsic, extrinsic, and in between.
TrendsinNeurosciences,27(7), 384-391.
Weddell, G., Taylor, D. A., & Williams, C. M. (1955). Studies on the innerva-
tion of skin, III: The patterned arrangement of the spinal nerves to the
rabbit ear. Journal of Anatomy, 89, 317-342.
Complexty, Informaton, and Robustness
Chapter.VII
Complexity,.Information,.
and.Robustness:
The.Role.of.Information.
Barriers.in.Boolean.Networks
Kurt A. Rchardson, ISCE Research, USA
Abstract
In this supposed information age, a high premium is put on the widespread
availability of information. Access to as much information as possible is often
cited as key to the making of effective decisions. While it would be foolish
todenythecentralrolethatinformationanditsfowhasineffectivedeci-
sion-making processes, this chapter explores the equally important role of
barrierstoinformationfowsintherobustnessofcomplex systems. The
analysis demonstrates that (for simple Boolean networks at least) a complex
systemsabilitytoflterout(i.e.,block)certaininformationfowsisessential
if it is not to be beholden to every external signal. The reduction of informa-
tion is as important as the availability of information.
Rchardson
Introduction
In the information age, the importance of having unfettered access to infor-
mation is regarded as essentialalmost a rightin an open society. It is
perhaps obvious that acting with the beneft of (appropriate) information to
hand results in better actions (i.e., actions that are more likely to achieve
desired ends), than acting without information (although incidents of infor-
mation overload and paralysis by (over) analysis are common). From a
complex systems perspective there are a variety of questions/issues concerning
information, and its near cousin knowledge, that can be usefully explored. For
instance, what is the relationship between information and knowledge? What
is the relationship between information, derived knowledge, and objective
reality? What information is necessary within a particular context in order to
make the best choice? How can we distinguish between relevant information
and irrelevant information in a given context? Is information regarding the
current/past state of a system suffcient to understand its future? Complexity
thinking offers new mental apparatus and tools to consider these questions,
often leading to understanding that deviates signifcantly from (but not neces-
sarily exclusive of) the prevailing wisdom of the mechanistic/reductionistic
paradigms. These questions may seem rather philosophical in nature, but
with a deeper appreciation of the nature of information and the role it plays
in complex systems and networks, we can begin to design more effective
and effcient systems to facilitate (rather than merely manage) information
creation, maintenance, and diffusion.
The science of networks has experienced somewhat of a renaissance in
recent years with the discovery of particular network topologies, or ar-
chitectures, in natural and human systems. These topologies include both
small-world (Watts & Strogatz, 1998) and scale-free networks (Barabsi
& Albert, 1999). Barabsi (2001) has shown that the World Wide Web has
a scale-free architecture, which essentially means that it has relatively few
high-connected nodes (i.e., nodes containing many inputs and outputs) and
relatively many low-connected nodes (i.e., nodes containing few inputs and
outputs). However, there are signifcant limitations to network representa-
tions of real world complex systems. Barabsi (2001) himself says that The
advances [in network theory] represent only the tip of the iceberg. Networks
represent the architecture of complexity. But to fully understand complex
systems, we need to move beyond this architecture and uncover the laws
that govern the underlying dynamical processes, such as Internet traffc
Boolean network modeling is a relatively simple method to facilitate this
move beyond mere architecture. Although the networks considered herein
have random topologies (rather than scale-free or small-world) the model-
ing of such simple dynamical networks allows researchers to explore, albeit
in a limited fashion, the emergent underlying dynamical processes of real
systems such as the Internet.
In this chapter, I would like to examine one particular aspect of complex
dynamical networks: How barriers to information and its fow are essential
in the maintenance of a coherent functioning organization. My analysis
will be necessarily limited. A very specifc type of complex system will be
employed to explore the problem, namely, Boolean networks. And, a rather
narrow type of information will be utilized: a form that can be recognized
by such networks. Despite these and other limitations, the resulting analysis
has applicability to more realistic networks, such as human organizations,
and useful lessons can be gleaned from such an approach. We shall see that
such an approach may offer the possibility of complementing mechanistic
approaches to the design of information systemsin which the most im-
portant characteristics are engineeredwith approaches that allow certain
characteristics to emerge from the interaction of the information systems
components. The chapter begins with an introduction to Boolean networks
and certain properties that are relevant to the analysis herein.
Boolean.Networks:.Their.Structure.and.
Their.Dynamics
Given the vast number of papers already written on both the topology and
dynamics of such Boolean networks, there is no need to go into too much
detail here. The interested reader is encouraged to look at Kauffman (1993)
for his application of Boolean networks to the problem of modeling genetic
regulatory networks. A short online tutorial is offered by Lucas (2006), but
the basics are provided herein.
A Boolean network, which is a particularly simple information system, is
described by a set of binary gates S
i
={0,1} interacting with each other via
certain logical rules and evolving discretely in time. The binary gate represents
the on/off state of the atoms (or, agents) that the network is presumed
to represent. So, in a genetic network, for example, the gate represents the
0 Rchardson
state of a particular gene. The logical interaction between these gates would,
in this case, represent the interaction between different genes. The state of
a gate at the next instant of time (t+1) is determined by the k inputs or con-
nectivity
m
i
j
S
(m=1,2,,k) at time t and the transition function f
i
associated
with site i:
( )
1 2
( 1) ( ), ( ), , ( )
k
i i i
j j j
i i
S t f S t S t S t + =
There are 2
k
possible combinations of k inputs, each combination producing
an output of 0 or 1. Therefore, we have
2
2
k
possible transition functions, or
rules, for k inputs. For example, if we consider a simple network in which
each node contains 2 inputs, there are 2
2
=4 possible combinations of input
- 00, 01, 10, 11 - and each node can respond to this set of input combina-
tions in one of 2
2
2
=16 different ways. Figure 1 illustrates a simple example.
The fgure shows not only an example network (a), but also its space-time
evolution (b) and its phase space attractors (c) which will be discussed next.
Figure 1d shows the sixteen different Boolean functions that can be created
from two inputs.
The state, or confguration, space for such a network contains 2
N
unique states,
where N is the size of the network (so for the simple example shown, state
space contains 64 states). Because state space is limited in size, as the system
is stepped forward it will eventually revisit a state it has visited previously.
Combine this with the fact that from any state the next state is unique (although
multiple states may lead to the same state, only one state follows from any
state), then any Boolean network will eventually follow a cycle in state space.
As a result, state space (or phase space) is connected in a non-trivial way,
often containing multiple attractors each surrounded by fbrous branches of
states, known as transients. Figure 1c shows the state space attractors, and
their associated basins, for the particular Boolean network shown in Figure
1a. The transition functions for each node were chosen randomly, but did
not include the constant rules 0 (0000), or 15 (1111) as these force the node
to be input-independent nodes with constant transition functions that do not
change from their initial state.
Figure 1c shows the two attractors for this particular example: one period-
3 attractor basin containing 56 states, and a single period-2 attractor basin
containing the remaining eight states (a total of 2
5
=64 states). In this example,
Figure 1. A simple Boolean network (a) shows the set-up of a particular
networkwithN=6.Eachnodeisnumbered0thru5.Thebinarysequences
in between the square brackets represent the Boolean function for each node.
Notethatnodes2and5areconnectedtothemselves.(b)showsthespace-
time diagram for the evolving network. Each column shows the evolution of
a particular nodes, and each row shows the overall network state at a par-
ticulartime-step.Inthiscasethenetworkwasseededatt=0witharandom
sequence. After an initial settling down period, the sequence converges of
aperiod-3cycle.(c)showsthetwostatespaceattractorsthatcharacterize
thisparticularnetwork:aperiod-3(p3)andaperiod-2(p2).(d)showsall
the sixteen Boolean logic functions that can be constructed from two inputs
depicted in the standard way with both look-up tables and machine state
diagrams. In this study the two rules CONTR. (contradiction) and TAUT.
(tautology) are excluded.
Rchardson
only fve states actually lie on attractors and the remaining 59 lie on the tran-
sient branches that lead to these four cycle states. When a Boolean network
is simulated, it is usually seeded with an initial random confguration. As the
network is stepped forward one time-step at a time, there is an initial settling
down period (sometimes referred to as the relaxation time) before the system
converges onto a particular attractor. There are often many different routes to
a particular attractor, and these are represented by the branching structures
(or, transient branches) shown in Figure 1c. This relaxation period might be
compared to the settling down period in human systems that occurs when a
new strategy is implemented and people take time to fnd their place in the
new strategy before the overall system stabilizes in one particular direction
or another. Of course, this relatively stable period may only be temporary as
adjustments to strategy are made in response to changing requirements often
triggered by particular environmental conditions.
Figure 2a is another way of visualizing how state space is connected. Each
shade of grey represents one of the two attractor basins that together char-
acterize state space. The fgure comprises a grid with state 000000 at the
uppermost left position and state 111111 at the lowermost right position. It
shows, for example, that the 8 states that converge on the period-2 attractor
are distributed in four clusters, each containing two states. This particular
representation is called a destination map, or phase space portrait, and we
can see that even for small Boolean networks, the connectivity of state space
is non-trivial. Figure 2b is a destination map for an N=15, k=2 Boolean
network (whose state space is characterized by four attractors: 2p2, 2p3), to
illustrate how quickly the complexity of state space connectivity, or topology,
increases as network size, N, increases. The boundaries between different
shades are known as separatrices, and crossing a separatrix is equivalent to
a move from one attractor basin into another (adjacent) attractor basin (i.e.,
if the system is pushed across a separatrix then there will be a qualitative
change in dynamic behaviorthis is also known as bifurcation).
It is all well and good to have techniques to visualize the dynamics that results
from the Boolean network structure (and rules). But how are a networks
structure and its dynamics related? It is in the consideration of this question
that we can begin to explore the role of information fows in such systems.
The potentially complex dynamics that arise from the apparently simple
underlying structure is the result of a number of interacting structural feed-
back loops. In the Boolean network modeled to create the images in Figure
1, there are seven interacting structural feedback loops: two period-1, two
period-2, two period-3, and one period-5, or 2P1, 2P2, 2P3, 1P5 for short.
The P5 structural loop in this case is: Node 0 4 3 2 1 0. It is
the fow (and transformation) of information around such structural loops,
and the interactions between different loops (or, inter-loop connectivity)
that results in the two state space attractors in Figure 1c; 1p1, 1p3 for short
(here I use P for structural, or architectural, loops and p for state, or
phase space loops, or cycles). The number of structural P-loops increases
exponentially (on average) as N increases, for example, whereas the number
of state space p-loops increases (on average) in proportion to N (not
N
as
Kauffman (1993) reported, which was found to be the result of sampling
bias, Bilke & Sjunnesson (2002).
For networks containing only a few feedback loops, it is possible to develop
an algebra that can relate P-space to p-space, that is, it is possible to construct
some very simple and robust rules that relate a networks architecture to its
function much in the same way as it is rather trivial (when compared to com-
Figure2.Thedestinationmapsfor(a)thenetworkdepictedinFigure1a
(N=6, k=2), and (b) a particular N=15, k=3 network (the grid showing
individual states is not included for clarity)
(a) (b)
Rchardson
plex nonlinear dynamical systems) to build serial computers with particular
functionality. However, as the number of interacting p-loops increases this
particular problem becomes intractable very quickly indeed, and the devel-
opment of a linking algebra (i.e., rules) utterly impractical.
Sometimes the interaction of a networks P-loops will result in state space
(p-space) collapsing to a single period-1 attractor in which every point in state
space eventually reaches the same state. Sometimes, a single p-loop will result
whose period is much larger than the size of the networksuch attractors are
called quasi-chaotic (and can be used as very effcient and effective random
number generators). Most often multiple attractors of varying periods result,
which are distributed throughout state, space in complex ways.
Before moving on to consider the robustness of Boolean networks, which
will then allow us to consider the role information barriers play in network
dynamics, it should be noted that state space, or phase space, can also be
considered to be functional space. The different attractors that emerge from
a networks dynamics represent the networks different functions. For ex-
ample, in Kauffmans analogy with genetic regulatory networks, the network
structure represents the genotype, whereas the different phase space attractors
represent the resulting phenotype, each attractor representing a different cell
type. Another example might be in team dynamics in which the modes of
forming, storming, norming, and performing are each represented in state
space by a different attractor. A further example might be in computer network
operations. Here again, different operational modes, including both stable
and unstable (e.g., a crash) modes can be considered as different state space
attractors. An appreciation of a systems state space structure tells us about
the different responses a system will have to a variety of external perturba-
tions (i.e., it tells us which contextual archetypes the system sees and is
able to respond to).
Dynamical.Robustness
The dynamical robustness of networks is concerned with how stable a par-
ticular network confguration, or operational mode, is under the infuence
of small external signals. Most information systems will exhibit a range of
qualitatively different modes of operation. Only some of those modes will
be desirable, and others (like system wide crashes) will be undesirable. An
understanding of a systems dynamical robustness provides insights into which
modes are most likely (i.e., which ones dominate state space), and how easily
the systems behavior can be pusheddeliberately or notinto a differ-
ent mode. In Boolean networks, we can assess this measure by disturbing an
initial confguration (by fipping a single bit/input, i.e., reversing the state of
one of the systems nodes) and observing which attractor basin the network
then falls into (after some relaxation period). If it is the same attractor that
follows from the unperturbed (system) state then the state is qualitatively
stable when perturbed in this way. An average for a particular state (or, system
confguration) is obtained by perturbing each bit in the system state, or each
node, and dividing the number of times the same attractor is followed by the
network size. For a totally unstable state the robustness score would be 0,
and for a totally stable state the robustness score would be 1. The dynamical
robustness of the entire network is simply the average robustness of every
system state in state space. This measure provides additional information
concerning how state space is connected in addition to knowing the number
of cyclic attractors, their periods, and their weights (i.e., the volume of state
space they occupy). By way of example, the average dynamical robustness
of the network depicted in Figure 1a is 0.875 which means the network is
qualitatively insensitive to 87.5% of all possible 1-bit perturbations.
More on Structure and Dynamics: Walls of
Constancy,.Dynamics.Cores,.and.Modularization
As we have come to expect in complex systems research, there is always
more to the story than what at frst meets the eye. This is also the case for
the relationship between network structure and dynamics. Although infor-
mation fows (and is transformed) around the various structural networks,
certain logical couplings emerge that force particular nodes into one state
or another, keeping them in that state for as long as the network is simulated.
In other words, once the network is initiated/seeded and run forward, after
some seemingly arbitrary transient period (the relaxation time) some nodes
cease to allow information to pass. These fxed. or frozen nodes effec-
tively disengage all structural feedback loops that include themalthough
these structural loops still exist, they are no longer able to cycle information
continuously around them. As such we can refer to them as non-conserv-
ing information loops; information that is placed onto such loops will fow
around the loop until a frozen node is reached, after which the information
Rchardson
Figure3.Theprocessofmodularization in complex networks
fow is blocked. A number of such nodes can form walls of constancy through
which no information can pass, and effectively divide the network up into
non-interacting sub-networks (i.e., the network is modularized). This process
is illustrated in Figure 3.
Nodes/Loop.Types
The process is actually more complicated than what is depicted in Figure 3.
When the occurrence of frozen nodes is examined more closely we fnd
that some nodes are not frozen in all the different operational modes (attrac-
tors). As such, we can distinguish between four different types of informa-
tion loop:
1.. Type1: Those that contain frozen nodes (that are frozen to the same
state, 0 or 1) in all modes of operation (attractors);
2.. Type.2:.Those that contain the frozen nodes in all modes of operation, but
the frozen state may differ from one mode of operation to another;
3.. Type.3:.Those that contain frozen nodes in some modes of operation,
but may be active (i.e., contain no frozen nodes) in other modesthese
act as non-conserving loops in some contexts, but as conserving loops
in other contexts);
4.. Type.4:.Those that contain no frozen nodes in any modes of operation
(i.e., information conserving loops).
Given that it is the existence of frozen nodes, or not, that determines whether
an information loop is Type 1, 2, 3, or 4, we can also label nodes as Type 1,
2, 3, or 4 also. Throughout the rest of this chapter when we talk of frozen
nodes, or non-conserving information loops, we are only concerned with
Type 1 as these do not contribute to the gross characteristics of state space
at all. But before moving on to consider the role of Type 1 nodes/loops at
length Id like to briefy mention the role of Types 2 and 3.
Assuming a particular network contains all types of nodes/loops, when the
system is following different attractors, different sets of nodes/loops will be
active. As we shall see below, only the dynamic core (which comprises
all types except Type 1 nodes/loops) contributes to the number and period
Rchardson
of attractors in state space. However, the allocation of node/loop type is a
global one in that we must consider the behavior of all nodes and loops in
all modes of operation. If we consider the behavior of all nodes and loops in
only one particular mode then each node becomes either frozen or not, and
each loop becomes either conserving or non-conserving (i.e., Types 2 and
3 are effectively either Types 1 or 4 at the local attractor level). As a result,
there exists another level of modularization. The modularization process pre-
viously discussed is the result of the emergence of only Type 1 nodes/loops
and as such the modules identifed in this way are the largest in size, but the
fewest in number (because every operational modeattractor typeis con-
sidered). However, for particular modes, there may be additional nodes that
are frozen (and therefore a greater number of non-conserving loops) and so
certain modules may actually divide (modularize) further. In this localized
scenario, although the static structure of the network (i.e., the pre-simulated
structure) is the same in all modes of operation, the emergent dynamic struc-
ture is (potentially) different for each state space structure (attractor). Allen
(2001) introduced the term structural attractor, which seems to be the same
as the emergent dynamic structure discussed here. The relationship between
structural attractors and state space attractors is complex (as one might expect).
For example, the same structural attractors may account for different state
space attractors (although in the case of Boolean networks I would expect
those different state space attractors to be the same in most respects, such
as period for example).
I will not explore the role of Type 3 and Type 4 nodes/loops further herein,
but I suspect that the key to understanding the dynamics of (discrete) complex
networks at both the global (state space wide) and local (for particular state
space attractors) levels is through an understanding of how these different
types form and interact.
The.Emergence.of.a.Dynamic.Core
The identifcation of frozen nodes is non-trivial (and even non-possible)
before the networks are simulated. Although the net effect is the same as as-
sociating a constant transition function (i.e., contradiction or tautology) with
a particular node, the effect emerges from the dynamic interaction of the
structural feedback loops (and is often dependent on initial conditions). It is
a rare case indeed that these interactions can be untangled and the emergent
frozen nodes identifed analytically beforehand.
As previously mentioned, these frozen nodes (Type 1) do not contribute
to the qualitative structure of state space, or network function. A Boolean
network characterized by a 1p3, 1p2 state space, like the one considered pre-
viously, will still be characterized by a 1p3, 1p2 state space after the (Type
1) frozen nodes are removed. In this sense, they dont appear to contribute
the networks functionthey do block the fow of information, but from this
perspective have no functional role. Another way of saying this is that, if we
are only concerned with maintaining the qualitative structure of state space
(i.e., the gross functional characteristics of a particular network) then we need
only concern ourselves with information conserving loops; those structural
feedback loops that allow information to fow freely around the network.
This feature is illustrated in Figure 4 where the state space properties of a
particular Boolean network, and the same network after the non-conserving
information loops have been removed (i.e., the networks reduced form),
are compared. Although the transient structure (and basin weight) is clearly
quite different for the two networks, they both have the same qualitative
phase space structure: 2p4.
The process employed to identify and remove the non-conserving loops is
detailed in Richardson (2005a). As network size increases, it becomes increas-
ingly diffcult to determine a networks state space structure. As such reduction
techniques are not only essential in facilitating an accurate determination,
but also in research that attempts to develop a thorough understanding of the
relationship between network structure and network dynamics (as, already
mentioned, only information conserving structural loops contribute to the
networks gross functional characteristics). Reduction techniques basically
sample state space and identify, which nodes are frozen to the same state (0
or 1) in all attractors (see the previous defnition of Type 1 nodes). These
nodes, and nodes that have no outgoing connections (and any connections
associated with these nodes), are removed from the network description to
form a reduced version. Another way of saying this is that all non-con-
serving information loops are identifed and deactivated. What remains after
this reduction process is known as the networks dynamic core. A dynamic
core contains only information conserving loops. The majority of (random)
Boolean networks comprise a dynamic core (which may be modularized)
plus additional nodes and connections that do not contribute to the asymp-
totic dynamics of the network (i.e., the qualitative structure of state space).
A full description of a systems dynamic core is the smallest description of
the system that still contains the essence of what that systems function is:
all other structure is superfuous.
00 Rchardson
Figure4.Anexampleof(a)aBooleannetwork,and(b)itsreducedform.
The nodes, which are made up of two discs feedback onto themselves. The
connectivity and transition function lists at the side of each network repre-
sentation are included for those readers familiar with Boolean networks.
The graphics below each network representation show the attractor basins
foreachnetwork.Thestatespacesofbothnetworkscontaintwoperiod-4
attractors, although it is clear that the basin sizes (i.e., the number of states
they each contain) are quite different. The window located at the top-middle
ofthefgureillustrateshowdifferentrulesareapplied(inthetwo-inputcase).
The different combinations of the two-input, A and B, are mapped to outputs,
O.Intheexamplegiven(whichisforrule[0110])theruletableshows
that if both A and B or zero or one, then the next state (O) of the node that
hasAandBasinputs,willbe0.ThisistheessenceofhowBooleannetworks
are stepped forward in time.
(a) (b)
Complexty, Informaton, and Robustness 0
If we consider a clock: at the heart of most clocks found in the home there is
some kind of time-keeping mechanism and another mechanism that represents
the time kept to observers of that clock. Of course, these mechanisms are
often surrounding by fancy cases and other sub-systems designed for aesthetic
reasons, or to protect the main mechanisms from external infuence. We could
regard the central time-keeping mechanisms as this particular systems dy-
namic core, with the other frills and fourishes largely being irrelevant if our
concern is with the time-keeping functionality of the clock. It should be noted
though that in this example, the existence of modules, sub-systems, barriers,
etc. are engineered into the system and that the dynamic core is trivial to
identify (once a particular functional perspective is chosen). In the Boolean
networks considered in this study, the existence of modules, sub-systems,
barriers, etc. emerges through the nonlinear interaction of the various system
components, and so details of their existence cannot (normally) be determined
beforehand. It is likely that in real world networks such as the World Wide
Web (WWW) some of the structure will be hard-wired (designed) into the
network, and other structures (which may not be easily identifed) emerge.
Another way of thinking about a complex networks dynamic core is in the
process of modeling itself. Complex systems are incompressible (e.g., Cil-
liers, 1998; Richardson, 2005b). This statement simply asserts that the only
complete model of a particular complex system is the system itself
1
. This is,
however, quite impractical, especially when considering systems such as the
WWW or a natural ecology. Instead of regarding our models as efforts to
represent a system in its entirety we can say, using the language developed
above, that the aim of any modeling process is to fnd an adequate representa-
tion of a systems dynamic core (i.e., those processes that are central to the
functionality of the system).
The.Role.of.Non-Conserving.(Structural).
Information.Loops
As the gross state space characteristics of a Boolean network and its dynamic
core (or, reduced form) are the same (i.e., in this sense, they are functionally
equivalent), it is tempting to conclude that (Type 1) non-conserving infor-
mation loopsinformation barriersare irrelevant. If this were the case then
it might be used to support the widespread removal of such dead wood
0 Rchardson
from complex information systems (e.g., human organizations). The history
of science is littered with examples of theories which once regarded such
and such a phenomena as irrelevant, or waste, only to discover later on
that it plays a very important role indeed. The growing realization that junk
DNA (DNA being probably the most intricate information system known
to us, other than the brain perhaps) is not actually junk is one such example.
What is often found is that a change of perspective leads to a changed as-
sessment. Such a reframing leads to a different assessment of non-conserving
information loops. Our limited concern, thus far herein, with maintaining a
qualitatively equivalent state space structure in the belief that a functionally
equivalent network is created, supports the assessment of non-conserving
information loops as junk. However, this assessment is wrong when ex-
plored from a different angle.
There are at least two roles that non-conserving information loops play in
random Boolean networks:
1. The process of modularization, and;
2. The maximization of robustness.
Modularization.in.Boolean.Networks
We have already briefy discussed the process of modularization. This pro-
cess, which we might label as an example of horizontal emergence (Sulis,
2005) was frst reported by Bastolla and Parisi (1998). It was argued that the
spontaneous emergence of dynamically disconnected modules is key to un-
derstanding the complex (as opposed to ordered and quasi-chaotic) behavior
of complex networks. So, the role of non-conserving information loops is to
limit the networks dynamics so that it does not become overly complex, and
eventually quasi-chaotic (which is essentially random in this scenario: when
you have a network with a high-period attractor of say 10
20
which is not
hard to obtainit scores very well indeed against all tests for randomness.
One such example is the lagged Fibonacci random number generator).
In Boolean networks, the resulting modules are independent of each other,
so the result of modularization, is a collection of completely separate sub-
systems. This independency is different from what we see in nature, but the
attempt to understand natural complex systems as integrations of partially
independent and interacting modules is arguably a dominant theme in the life
sciences, cognitive science, and computer science (see, for example, Callebaut
& Rasskin-Gutman, 2005). It is likely that some form of non-conserving,
or perhaps limiting, information loop structure plays an important role
in real world modularization processes. Another way of expressing this is
organization is the result of limitinginformationfow.
The concept of modularization (which is an emergent phenomenon) appears to
be similar to Simons (1962) concept of near decomposability. In his seminal
paper The Architecture of Complexity, Simon developed his theory of nearly
decomposable systems, in which the interactions among the subsystems are
weak, but not negligible (p. 474). Simon goes on to say:
At least some kind of hierarchic systems can be approximated successfully
asnearlydecomposablesystems.Themaintheoreticalfndingsfromtheap-
proach can be summed up in two propositions: (a) in a nearly decomposable
system, the short run behavior of each of the component subsystems is ap-
proximately independent of the short-run behavior of the other components;
(b) in the long run, the behavior of any one of the components depends in
onlyanaggregatewayonthebehavioroftheothercomponents(p.474).
The process of modularization previously described is slightly different, but
complementary. Near decomposability suggests that complex systems evolve
such that weakly interacting subsystems emerge that, on suffciently short
timescales, can be considered as independent from each other (although it says
little about the dynamic processes that lead to this state). This phenomenon is
clearly observed in Boolean networks, except that in these particular complex
systems, the emergent subsystems (modules) are completely independent of
each otherinformation does not fow from one subsystem to another. If
there were no communications with the systems environment then we could
simply consider these independent subsystems in isolation. However, we shall
see in the following discussions on dynamical robustness that the padding
between these subsystemsthe (Type 1) non-conserving information loops,
or frozen nodesplays an important role in the overall systems (the network
of subsystems) dynamics in the face of external interference. (It should be
noted that, although the emergent modules discussed here are independent,
in larger modules sub-modules may emerge that can indeed be viewed as
weakly interacting subsystems. Near decomposability and modularization
are related, but not in a trivial way.)
0 Rchardson
Dynamic.Robustness.of.Complex.Networks
Dynamic robustness was previously defned in a slightly different way, as
the stability of a networks qualitative behavior in the face of small external
signals. In this section, we will consider the dynamic robustness of an en-
semble of random Boolean networks and their associated reduced form to
assess any difference between the two.
To do this comparison, the following experiment was performed. 10
6
random
Boolean networks with N=15 and k =2 (with random connections and random
transition functions, excluding the two constant functions) were constructed.
For each network, its dynamic core was determined using the method detailed
in Richardson (2005a). The average dynamical robustness was calculated for
both the (unreduced) networks and their associated dynamic cores. The data
from this experiment is presented in Figure 5, which shows the relationship
between the unreduced and reduced dynamic robustness for the 10
6
networks
considered. The black points shows the average value of dynamic core (or,
reduced) robustness for various values of unreduced robustness. On average,
the dynamic robustness of the reduced networks is typically of the order of
20% less than their parent (unreduced) networks. Of course, the difference
for particular networks is dependent on specifc contextual factors, such as
the number of non-conserving information loops in the (unreduced) networks
(the extent of the dynamic core, in other words). This strongly suggests
that the reduced networks are rather more sensitive to external signals than
the unreduced networks. In some instances the robustness of the reduced
network is actually zero meaning that any external signal whatsoever will
result in qualitative change. This generally only occurs in networks, which
have small dynamic core sizes compared to the network size. What is also
interesting, however, is that sometimes the reduced network is actually more
robust than the unreduced network. This is a little surprising perhaps, but not
when we take into account the complex connectivity of phase space for these
networks. This effect is observed in cases when there is signifcant change in
the relative attractor basin weights as a result of the reduction process and/or
a relative increase in the orderliness of state space.
In complex systems research it is important to consider the process of av-
eraging data (and adding error bars extending a certain number of standard
deviations). The data in Figure 5a is clearly multi-modal and as such, one
has to take care in interpreting the physical meaning of the average. The
different shades of grey in Figure 5 indicate the number of data points that
Figure 5. Data histograms of the dynamical robustness data collected. (a)
all data, with black points showing the overall average, (b) data associated
withadynamiccoresizeof15,(c)14,(d)13,(e)12,(f)11,(g)10,(h)9,(i)
8,(j)7,(k)6,(l)5,and(m)4.Datafordynamiccoresizesof3,2,and1are
not included
0 Rchardson
fall into a particular data bin (the data is rounded down to 2 decimal places
leading to 100 data bins). So the darkest grey regions contain many data
points (>a
10
where a is the relative frequency of the data, a =2.4 in this case)
and the lightest grey regions contain only one (or a
0
) data point. If we had
blindly calculated the average values and attempted to interpret its mean-
ing, we would have missed the importance of dynamic core size completely.
Figures 8b-8m show only the data for particular sizes of dynamic core. This
helps considerably in understanding the detailed structure of Figure 8a. The
various diagonal modal peeks relate to networks with different dynamic
core sizes, and the different horizontal structures correlate with networks
containing smaller dynamic cores which can have only limited values of
dynamic robustness (i.e., as the size of the dynamic core decreases the data
Figure 6. A data histogram showing the relationship between the number
of structural feedback loops in the unreduced networks, and the number of
active structural loops in their dynamic cores. The black points represent
the average number of structural loops in dynamic core.
appears more discrete and as the core size increases the data appears more
continuous).
Further analysis was performed to confrm the relationship between network
structure, dynamic core structure, and state space structure. This included
comparing the number of structural feedback loops in the overall network
to the number of (information conserving only) loops in networks dynamic
core. Figure 6 shows the data for all 10
6
(N=15, k=2) networks studied with
all dynamic core sizes superimposed on each other. The data indicates that
on average the dynamic core of a network has between 30% and 60% fewer
structural feedback loops; all of them information conserving loops. On
average only 42% of all structural feedback loops contribute to the (global)
functionality of a particular network. It should of course be noted though
that this proportion is strongly dependent upon dynamic core size. In the
next section, we shall consider the implications of this in terms of state space
characteristics and dynamic robustness.
State.Space.Compression.and.Robustness
In Boolean networks, each additional node doubles the size of state space.
In fact, in any discrete system (or any system that can be approximated as
discrete), when an additional node is added, the size of state space is increased
by a factor equal to the total number of states that the additional node can
adopt. So in a human system, like a team for example, the addition of an extra
member will increase state space by a factor equal to the number of different
responses that the additional member exhibits when operating in that particular
system. This neednt result in the appearance of new state space structures
(attractors), but the volume of state space is greatly increased. So even if
frozen nodes contribute nothing to the qualitative structure of state space
(which I shall now refer to as frstorderfunctionality), they at least increase
the size of phase space. As an example, the phase space of an N=20 network
is 1024 times larger than an N=10 network. Thus, node removal signifcantly
reduces the size of phase space. As such, the chances that a small external
signal will inadvertently target a sensitive area of state space, i.e., an area
close to a separatrix, therefore pushing the network toward a different attrac-
tor, are signifcantly increased: a kind of qualitative chaos. This explains why
we see the robustness tending to decrease when non-conserving information
0 Rchardson
loops are removed: the (emergent) buffer between the system and the outside
worldthe systems environmenthas been removed.
Prigogine said that self-organization requires a container (self-contained-
organization). Non-conserving information loops function as a kind of
container. So it seems that, although non-conserving information loops do
not contribute to the long term behavior of a particular network, these same
loops play a central role as far as the dynamical stability is concerned. Any
management team tempted to remove 80% of their organization in the hope
of still achieving 80% of their yearly profts (which is sometimes how the
80:20 principle in general systems theory is interpreted in practice) would
fnd that they had created an organization that had no protection whatsoever
to even the smallest disturbances from its environmentit would literally
be impossible to have a stable business.
It should be noted that the non-conserving information loops do not act as
impenetrable barriers to external signals (information). These loops simply
limit the penetration of the signals into the system. For example, in the case
of a modularized network, the products of incoming signals may, depending
on network connectivity, still be fed from a non-conserving information loop
onto information conserving loops for a particular network module. Once
the signals have penetrated a particular module, they cannot cross over into
other modules (as the only inter-modular connections are via non-conserv-
ing loops). As such, it would seem that non-conserving loops may play an
information distribution role.
It should also be noted that even though a particular signal may not cause
the system to jump into a different attractor basin, or bifurcate, it will still
push the system into a different state on the same basin. The affect of signals
that end-up on non-conserving information loops is certainly not nothing.
So, although the term information barriers is used herein, these barriers
are semi-permeable.
Balancing.Response.Strategies.and.
System.Robustness
Figure 7 shows a data histogram for the number of state space attractors vs.
(unreduced) dynamic robustness. The plot shows that the robustness decreases
rapidly initially as the number of state space attractors increases. Remem-
bering that the number of state space attractors can also be regarded as the
number of contextual archetypes that a system sees, we see that versatility
(used here to refer to the number of qualitatively different contexts a system
can respond to, or is sensitive to) comes at the cost of reduced dynamic ro-
bustness assuming the same resources are available. Considered in this way
robustness and versatility are two sides of the same coin. We would like for
our systems to be able to respond to a wide variety of environmental contexts
with minimal effort, but this also means that our systems might also be at
the mercy of any environmental change. A system with only one state space
attractor doesnt see its environment at all, as it has only one response in
all contexts, whereas a system with many phase space attractors sees too
muchthere is a price to pay for being too fexible.
This is only the case though for fxed resources, that is, given the same resources
a system with few state space attractors (modes of operation) will be more
robust than a system with a greater number. This is because, on average, the
system with more state space attractors will have a larger dynamic core and
so the buffering afforded by non-conserving loops will be less pronounced.
It is a trivial undertaking to increase phase space by adding nodes that have
Figure7.Adatahistogramshowingtherelationshipbetweenthenumberof
statespaceattractors(whichisalsorelatedtonetworkfexibility/versatility)
and the (unreduced) dynamical robustness. The black points represent the
average dynamic robustness for increasing numbers of phase space attrac-
tors (the greyscaling is scaled using a
n
,wherea=2.4andnisaninteger
between0and10inclusive).
0 Rchardson
inputs but no outputs, or leaf nodes (which is equivalent to adding con-
nected nodes that have a constant transition function)we might refer to
this as a frst-order strategy. This certainly increases the size of phase space
rapidly, and it is a trivial matter to calculate the effect such additions will
have on the networks dynamical robustness (i.e., the increased dynamical
robustness will not be an emergent property of the network). Increasing the
robustness of a network (without changing the number, period and weight of
phase space attractor basins) by adding connected nodes with non-constant
transition functions is a much harder proposition. This is because of the new
structural feedback loops created, and the great diffculty in determining
the emergent logical couplings that result in frozen nodes, which turn
structural information loops (that were initially conserving) into non-con-
serving loops. It is not clear at this point in the research whether there is any
preference between these two strategies to increase dynamical robustness,
although the emergent option is orders of magnitude harder to implement
than the other, and will likely change functionality.
One key difference between the straightforward frst-order strategy and the
more problematic (second-order) emergent strategy is that the frst-order
enhanced network would not be quantitatively sensitive to perturbations on
the extra (buffer) leaf nodes (i.e., not only would such perturbations not
lead to qualitative change (a crossing of a separatrix, or a bifurcation), but
the position on the attractor cycle that the system was on when the node was
perturbed (i.e., cycle phase) would also not be affected)there would be no
effect on dynamics whatsoever. This is because the perturbation signal (in-
coming information) would not penetrate the system further than the leaf
node, as by defnition it is not connected via any structural (non-conserving
or conserving) feedback loops: they really are impenetrable barriers to in-
formation. If the emergent strategy was successfully employed, the incom-
ing signals would penetrate and quite possibly (at most) change the cycle
phase (i.e., quantitative (but, again, not qualitative) change would occur).
Either one of these response traits may or may not be desirable depending
on response requirements.
Conclusion
From the analysis presented herein, it is clear that non-conserving informa-
tion loopsinformation barriersplay an important role in a networks
global dynamics: they are not expendable. They protect a network from
both quantitative and qualitative (quasi) chaos. Quantitative chaos is resisted
by the emergent creation of modulesthrough the process of modulariza-
tionwhich directly reduces the chances of state space being dominated by
the very long period attractors associated with quasi-randomness in Boolean
networks. Whereas qualitative chaosthe rapid jumping from one attrac-
tor basin to another in response to small external signalsis resisted by the
expansion of state space, which reduces the possibility of external signals
pushing a system across a state space separatrix (i.e., into another mode of
operation).
Much more effort will be needed to fully understand the design and operating
implications of this research for real world information systems. In one sense,
Boolean networks are at the heart of all technology based systemscomput-
ers are themselves very complicated Boolean networks. However, these sorts
of (complicated) Boolean networks do not usually have architectures that
contain the many nonlinear interacting feedback loops that the (complex)
Boolean networks discussed herein have. Complicated Boolean networks,
such as computers, generally process information serially, whereas complex
Boolean networks can process information in parallel. In fact, there is consid-
erable research being performed that considers complex Boolean networks as
powerful parallel computers (see for example, Mitchell, 1998). It may well
be that the sort of research presented here has little impact on the design of
engineered systems such as computers and technology-driven information
nodes as these are generally hard-wired not to exhibit emergent behavior. It is
in the area of soft-wired systems, perhaps, that this research may offer some
insight. For example, although the individual components of the Internet are
hard-wired, the way in which its global architecture is developing is through
emergent processes. Recent advances in network theory provide powerful
tools to understand the static structure (architecture) of complex networks
such as the WWW. But, equally powerful dynamical tools will be needed to
understand the dynamic structure of the WWW which is likely quite different.
There is clearly someway to go to fully understand what abstract concepts
such as non-conserving information loops, or frozen nodes, refer to in real
world information systems, but such research is beginning to provide us with
the language and tools to facilitate our understanding of the structural and
state space dynamics of a wide range of complex information systems like the
Internet, the WWW, or even large decision networks to mention just a few.
Another useful avenue of further research may come by considering the
activity of each node. In this chapter, we could have defned frozen
Rchardson
nodes as those with an activity of zero (i.e., their state does not change with
time). The remaining nodes will have different activity levels (e.g., a node
that changes state on every time step (a period-2 node) would have activity
1). The dynamic core structure of a particular network (containing non-in-
teracting modules) is found by removing all nodes with activity 0. However,
different structures will be revealed if the activity threshold for removal is
varied from 0 to 1. In this way, researchers would fnd further modularization,
although modules (sub-structure) found this way would not be completely
independent (the weakly interacting modules of Simons nearly decomposable
systems perhaps); systems of interacting sub-systems (with varying degrees
of interaction strength) within the same system would be found, which would
provide further insight into the relationship between emergent structure and
function. Furthermore, it is possible that such an activity threshold analysis
would be easier to apply to real world systems than the analysis performed
herein, which is only concerned with a zero activity threshold.
To conclude, I would suggest that barriers (both impenetrable and semi-
permeable) to information fow play a central role in the functioning of all
complex information systems. However, the implications (and meaning) of
this for real world systems is open to many different interpretations. At the
very least it suggests that barriers to information fow should be taken as
seriously as supports to information fow (although, paradoxically, a good
supporter is inherently a good barrier!). This may seem obvious given
the dual role of all types of boundarythat is, to keep in and keep out, or to
enable and disablebut this change in focus offers, perhaps, an interesting
counterpoint to the emerging cultural perspective that suggests that having
as much information as possible can only be a good thing.
Acknowledgment
Id like to thank Yasmin Merali and Bill McKelvey for useful suggestions
on how to improve the earlier drafts of this paper, and to three anonymous
reviewers. Thanks also go to Caroline, Alexander, and Albert for interesting
diversions.
References
Allen, P. M. (2001). What is complexity science? Knowledge of the limits
to knowledge.Emergence,3(1), 24-42.
Barabsi, A. L. (2001). The physics of the Web. Physics World. Retrieved
May 21, 2007, from http://physicsweb.org/articles/world/14/7/9
Barabsi, A. L., & Albert, E. (1999). Emergence of scaling in random net-
works. Science,286(15 Oct), 509-512.
Bastolla, U., & Parisi, G. (1998). The modular structure of Kauffman net-
works. Physica D, 115, 219-233.
Bilke, S., & Sjunnesson, F. (2002). Stability of the Kauffman model. Phys.
Rev. E, 65, 016129.
Cilliers, P. (1998). Complexity and postmodernism: Understanding complex
systems. London, UK: Routledge.
Callebaut, W., & Rasskin-Gutman, D. (2005). Modularity: Understanding
the development and evolution of natural complex systems. Cambridge,
MA: MIT Press.
Kauffman, S. A. (1993). The origins of order: Self-organization and selection
in evolution. New York: Oxford University Press.
Lucas, C. (2006). Boolean networksdynamic organisms. Retrieved May
21, 2007, from http://www.calresco.org/boolean.htm
Mitchell, M. (1998). Computation in cellular automata: A selected review. In
T. Gramss, S. Bornholdt, M. Gross, M. Mitchell, & T. Pellizzari (Eds.),
Nonstandard computation (pp. 95-140). Weinheim: VCH Verlagsge-
sellschaft. Note that cellular automata are essentially Boolean networks
with an ordered architecture rather than a random one.
Richardson, K. A. (2005a). Simplifying Boolean networks. Advances in
ComplexSystems,8(4), 365-381.
Richardson, K. A. (2005b). The hegemony of the physical sciences: An ex-
ploration in complexity thinking. Futures,37, 615-653.
Simon, H. (1962). The architecture of complexity. Proceedings of the Ameri-
can Philosophical Society, 106(6), 467-482. Reprinted in Richardson,
K. A., Snowden, D., Allen, P. M., & Goldstein, J . A. (Eds.) (2006).
Emergence:Complexity&Organization2005Annual,Volume7 (pp.
499-510). Mansfeld, MA: ISCE Publishing.
Rchardson
Sulis, W. H. (2005). Archetypal dynamical systems and semantic frames in
vertical and horizontal emergence. In K. A. Richardson, J . A. Goldstein,
P. M. Allen, & D. Snowden (Eds.), E:CO Annual Volume 6 (pp. 204-
216). Mansfeld, MA: ISCE Publishing.
networks. Nature,393(4), 440-442.
Endnote
1
Of course there is no such thing as a complete model. Models, by their
very nature, are always, incomplete. The importance of the concept of
incompressibility is to highlight that simplifcations of complex systems
can often lead (but not necessarily so) to understanding that is qualitatively
incomplete. Contrast this situation to that of complicated systems (i.e.,
systems that comprise many components that may also be related via
nonlinear interactions but without the connectivity profle (topology))
that would lead to the multiple nonlinear and interacting feedback loops
that distinguish complex systemsfor which simple representations
(models) can always be constructed that are qualitatively complete.
Emergent Specalzaton n Bologcally Inspred Collectve Behavor Systems
Chapter.VIII
Emergent.Specialization.in.
Biologically.Inspired.
Collective.Behavior.
Systems
G. S. Ntschke, Vrje Unverstet, The Netherlands
M. C. Schut, Vrje Unverstet, The Netherlands
A. E. Eben, Vrje Unverstet, The Netherlands
Abstract
Specialization is observable in many complex adaptive systems and is thought
by many to be a fundamental mechanism for achieving optimal effciency
within organizations operating within complex adaptive systems. This chapter
presents a survey and critique of collective behavior systems designed using
biologicallyinspiredprinciples.Specifcally,weareinterestedincollective
behavior systems where specialization emerges as a result of system dynamics
and where emergent specialization is used as a problem solver or means to
increase task performance. The chapter presents an argument for developing
Ntschke, Schut, & Eben
design methodologies and principles that facilitate emergent specialization
in collective behavior systems. Open problems of current research as well as
future research directions are highlighted for the purpose of encouraging the
development of such emergent specialization design methodologies.
Introduction
Specialization is observable in many complex adaptive systems
1
and is thought
by many to be a fundamental mechanism for achieving optimal effciency
within certain complex adaptive systems. In complex ecological communi-
ties, specializations have evolved over time as a means of diversifying the
community in order to adapt to the environment (Seligmann, 1999). Over
the course of evolutionary time, specialization in biological communities
has assumed both morphological (Wenseleers, Ratnieks, & Billen, 2003)
and behavioral forms (Bonabeau, Theraulaz, & Deneubourg, 1996). For ex-
ample, morphologically specialized castes have emerged in certain termite
colonies (Noirot & Pasteels, 1987), and honeybees dynamically adapt their
foraging behavior for pollen, nectar, and water as a function of individual
preference and colony demand (Calderone & Page, 1988). The consequence
of such specializations is that labor is effciently divided between special-
ized castes
2
and individuals for the beneft of accomplishing group tasks. In
such a sense, specialization can be viewed as an adaptive mechanism in a
complex adaptive system.
Many artifcial complex adaptive systems that exhibit collective behavior have
used design principles, which draw their inspiration from examples of spe-
cialization in nature. Such examples include complex ecological communities
such as social insect colonies (Bonabeau et al., 1996; Bonabeau, Sobkowski,
Theraulaz, & Deneubourg, 1997; Calderone et al., 1988; Noirot et al., 1987;
Seligmann, 1999; Wenseleers et al., 2003) biological neural networks (Baev,
1997), multi-cellular organisms (Hawthorne, 2001), economies of a nation,
companies, corporations, and other business organizations (Abdel-Rahman,
2001; Ng & Yang, 1997; Resnick, 1997). Such biologically inspired design
principles are especially prevalent in multi-robot (Potter, Meeden, & Schultz,
2001) swarm intelligence (Bonabeau, Dorigo, & Theraulaz, 1998) and arti-
fcial life systems (Nishimura & Takashi, 1997) where it is highly desirable
to replicate the success of biological collective behavior systems.
Suppositions.of.Specialization
Given empirical evidence offered by research in both biological collective
behavior systems, and biologically inspired artifcial collective behavior
systems
3
, two key observations can be stated.
Specialization that assumes either behavioral or morphological forms
is often present in biological systems that exhibit collective behavior.
In biological systems that exhibit collective behavior, specialization
is benefcial in that it increases the effciency of the system, or allows
collective behavior tasks to be solved that could not otherwise be solved
by individuals within the system.
Given these observations, one can formulate the assumption that specialization
is benefcial in biological inspired artifcial complex adaptive systems that are
designed to solve certain types of collective behavior tasks. Examples of such
types of collective behavior tasks are presented in section Collective Behavior
Tasks and Specialization. In order for this assumption to be proved, this chapter
proposes the need to develop emergent behavior design methodologies
4
. Such
methodologies would dictate design and engineering principles for creating
an artifcial complex adaptive system capable of solving collective behavior
tasks that require or beneft from specialization. Ideally, such methodologies
would result in the production of artifcial complex adaptive systems that
yield emergent yet desired forms of specialization. As in biological systems,
this emergent specialization could then be harnessed and used by the system
for the beneft of either increasing task performance, or solving certain col-
lective behavior tasks, that could not otherwise be solved.
Chapter.Goal.and.Motivation:.Specialization.as.a.Problem.
Solver
The chapters scope is a survey and critique of collective behavior systems
designed using biologically inspired design principles that use emergent
specialization to solve collective behavior tasks. Such design principles in-
clude self-organization, learning, and evolution (Brooks, 1990). This chapter
presents an argument for utilizing emergent behavioral specialization as a
problem solver in biologically inspired artifcial complex adaptive systems.
Such utilization would be advantageous given the numerous real world ap-
plications where specialization is benefcial. Examples of such applications
are presented in the section Collective Behavior Tasks and Specialization.
This chapters motivation is similar to that cited for the organic computing
research endeavor (Mller & Sick, 2006). Organic computing has recently
achieved some success in investigating the notion of defning and measur-
ing concepts such as emergence and self-organization in large distributed
complex adaptive systems. The key idea is to utilize emergent phenomena
for the beneft of solving tasks in organic computing systems. An organic
computing system is a technical system, which adapts dynamically to the
current conditions of its environment. It is self-organizing, self-confguring,
self-repairing, self-protecting, self-explaining, and context-aware (Mller et
al., 2006). Initial research in this area displays great promise, and includes
exploiting emergent functionality at the hardware level of visual micropro-
cessors for image recognition tasks (Komann & Fey, 2007), self-organizing,
and self-stabilizing role assignment in sensor and actuator networks (Weis,
Parzyjegla, J aeger, & Mhl, 2006), and self-organization of job scheduling
and distribution of jobs over nodes in a network (Trumler, Klaus, & Ungerer,
2006).
Chapter.Scope:.Behavioral.Specialization
Another important issue is which type of specialization
5
should be instituted
for the beneft of a collective behavior system. We have elected to only survey
research literature concerned with behavioral specialization. The decision to
adopt this focus was based on the discovery that with relatively few exceptions
(section: Types of Specialization) the majority of research concerning the use
of emergent specialization for improving task performance is restricted to
simulated systems. This is so, given the obvious engineering challenges and
inherent complexity of dynamically creating morphologically specialized ro-
bots and computer components, that represent effective solutions to emerging
challenges in a physical task environment (Parker & Nathan, 2006; Pfeifer,
Iida, & Gomez, 2006; Watson, Ficici, & Pollack, 1999b). Figure 1 presents the
scope of the chapter within the dimensions of emergent versus non-emergent
phenomena and behavioral versus morphological specialization.
Types.of.Specialization
Specialization in collective behavior systems has been studied from many
different perspectives (Bongard, 2000; Bryant & Miikkulainen, 2003; Blumen-
thal & Parker, 2004b; Campos, Theraulaz, Bonabeau, & Deneubourg, 2001;
Haynes & Sen, 1996b; Nolf et al., 2003b; Stone & Veloso, 2002; Whiteson,
Kohl, Miikkulainen, & Stone, 2003), and is thus often defned in accordance
with the goals of researchers conducting the study. Within collective behav-
ior literature, specialization is either studied as an emergent property of the
system, or is explicitly pre-programmed into the systems components. With
notable exceptions such as Funes, Orme, and Bonabeau (2003), there are few
examples of research that successfully specifes, a priori, what exactly the
behavior of system components should be, in order to produce a specifcally
desired, yet emergent collective behavior.
Figure 1. Types of specialization in biologically inspired collective behavior
systems.Thetopleft-handsidequadrantdefnesthescopeofthischapter.
Specifcally,adaptivesystemsthatuseheterogeneousorhomogenousdesign
approaches with the aim of deriving emergent behavioral specialization for
solving collective behavior tasks. See section: Types of Specialization for
details.
0 Ntschke, Schut, & Eben
Non-Emergent.Specialization
Non-emergent specialization is that which is explicitly pre-specifed to be
apart of the design of system components and global behavior of a system.
Such approaches are either static, or utilize learning algorithms so as to
ascertain which type of behavioral specialization, selected from a given set,
is most appropriate for solving a given task. Such approaches are useful for
solving collective behavior tasks that require specialization, where the degree
of specialization required can be suffciently described a priori (Arkin &
Balch, 1999; Balch, 2002a, 2002b).
Emergent.Specialization
Emergent specialization is that which emerges from the interaction of system
components in response to a dynamic task that requires varying degrees, or
different types of specialization, in order to effectively accomplish. Such
approaches have become popular in collective behavior task domains where
one does not know, a priori, the degree of specialization required to optimally
solve the given task (Gautrais, Theraulaz, Deneubourg, & Anderson, 2002;
Luke & Spector, 1996; Murciano & Millan, 1997a; Murciano, Millan, &
Zamora, 1997b; Potter et al., 2001; Stanley, Bryant, & Miikkulainen, 2005b;
Theraulaz, Bonabeau, & Deneubourg, 1998b; Waibel, Floreano, Magnenat,
& Keller, 2006). The section Heterogeneous vs. Homogenous Design of
Emergent Specialization elaborates upon such emergent specialization design
approaches.
Morphological.vs..Behavioral.Specialization
It is possible to further categorize specialization into two distinct classes:
morphological (Martinoli, Zhang, Prakash, Antonsson, & Olney, 2002; Zhang,
Martinoli, & Antonsson, 2003) and behavioral (Bonabeau et al., 1997; Li,
Martinoli, & Mostafa, 2002).
The term morphological specialization is applicable to situated and em-
bodied agents, operating in simulated or physical task environments, with
embodiment (sensors and actuators) structured so as to yield an advantage
in accomplishing the task (Watson et al., 1998a, 1999b, 2002). Examples of
morphological specialization include the evolution of optimal arrangements
of sensors and actuators in the design of simulated automobiles (Martinoli et
al., 2002; Zhang et al., 2003), evolution of agent morphologies and controllers
for various forms of motion in simulated environments (Sims, 2004), evolu-
tion of physical electric circuits for control (Thompson, Harvey, & Husbands,
1996), and evolving robot morphology for accomplishing different forms of
physical motion (Lipson & Pollack, 2000).
The term behavioral specialization is applicable to agents with behaviors that
are advantageous for accomplishing specifc types of tasks (Balch, 2002a,
2002b; Nolf & Floreano, 2000; Nolf & Parisi, 1997). Examples of behav-
ioral specialization include the use of machine learning methods that activate
certain behaviors with a particular frequency as a response to dynamically
arising tasks (Gautrais et al., 2002).
Collective.Behavior.Methods.for.Specialization
There is some agreement among researchers as to the methods for special-
ization that are appropriate for particular collective behavior tasks. Figure 2
illustrates a categorization of such methods, which are briefy detailed in the
following. The categories illustrated in Figure 2 are by no means exhaustive,
but rather several examples that have recently received particular research
attention.
Figure2.Collectivebehaviormethodsofspecialization.Seesection:Col-
lective Behavior Methods for Specialization for details.
Division.of.Labor.Methods
The use of behavioral threshold and division of labor methods have been
investigated within the context of ant-based (Deneubourg, Goss, Pasteels,
Fresneau, & Lachaud, 1987) and resource allocation (Bonabeau et al., 1997)
methods. Such methods typically utilize feedback signals given to agents of
the same caste (Kreiger & Billeter, 2000) in order to encourage the emer-
gence of specialization for a specifc task. Many variations of these methods
exist (Bonabeau & Theraulaz, 1999; Bonabeau et al., 1996, 1997, 1998;
Deneubourg et al., 1987; Robson & Traniello, 1999; Theraulaz, Gervet, &
Semenoff, 1991; Theraulaz, Goss, Gervet, & Deneubourg, 1991), including
those that use evolutionary algorithms (Tarapore, Floreano, & Keller, 2006;
Waibel et al., 2006), and reinforcement learning models (Murciano et al.,
1997a, Murciano et al., 1997b) in order to derive threshold values. The goal
of such models is typically to optimize global task performance. Such meth-
ods are appealing as their evolutionary dynamics and emergent properties
can usually be described with a mathematical representation and the results
of such models are thus typically amenable to a mathematical analysis (Wu,
Di, & Yang, 2003).
Mathematical,.Economic,.and.Game.Theory.Methods
Linear, non-linear, and dynamic methods based in mathematical, economic,
and game theory (Axelrod, 1984; Solow & Szmerekovsky, 2004) have many
applications for resource assignment problems in business. For example, the
maximum matching algorithm developed by Edmonds (1965) was designed
to determine the maximum number of people that can be assigned to tasks
in such a way that no person is assigned to more than one task. Thus, it is
assumed that each person specializes in performing at most one task. Such
methods are advantageous as results can be subject to a formal analysis.
However, they are limited by their abstract nature, and assume that the task
domain can be mathematically or otherwise formally represented.
Cooperative.Co-Evolution.Methods
Cooperative co-evolution methods have been implemented both in the context
of modifed genetic algorithms, for example, Cooperative Co-evolutionary
Genetic Algorithms (Potter & DeJ ong, 2000), and in the context of neuro-
evolution methods, for example, Enforced Sub-Populations (ESP) (Gomez,
1997). In both cases, the genotype space is decomposed into a set of sub-
populations, where each generation, the evolutionary process selects the best
performing genotype components from each sub-population so as to construct
a complete genotype as a solution. Decomposition of the genotype space
into sub-populations, genotype construction from multiple sub-populations,
and genotype to phenotype mapping depends upon the approach used. For
example, the ESP method encodes separate neurons as genotype components
to be distributed between sub-populations, where the composition of neurons
encodes a complete neural network. Advantages of such methods include
their versatility, and applicability to a broad range of complex, continuous,
and noisy task domains. Also, the representation of the genotype space as a
set of sub-populations provides a natural representation for many collective
behavior tasks, and often effectuates the derivation of specialized phenotypes.
A key disadvantage of such approaches is slow derivation of viable solu-
tions in complex task domains due to inherently large search spaces. Also,
the genotype representations that produce desired results can typically not
be easily interpreted.
Reinforcement.Learning.Methods
There exists a certain class of reinforcement learning methods that provide
periodic feedback signals to agent groups attempting to accomplish a collec-
tive behavior task (Sutton & Barto, 1998). A reinforcement signal is either
local or global. Local reinforcement signals are calculated by, and given to
a single agent, or a caste, upon task accomplishment. Global reinforcement
signals are calculated by and given to the entire agent group at the end of
a reinforcement learning trial (Li, Martinoli, & Yaser, 2004). The main
advantage of reinforcement learning approaches is that agents are able to
effectively operate in complex and noisy environments, with incomplete
information. However, approaches that utilize only a global reinforcement
signal, do not typically effectuate specialization in the group, even if task
performance could be increased with specialized agents (Li et al., 2002, 2004).
Approaches that utilize local reinforcement signals have been demonstrated
as being appropriate for deriving specialized agents (Li et al., 2004), however
such approaches suffer from the credit assignment problem (Grefenstette,
1995; Sutton et al., 1998), which potentially leads to sub-optimal collective
behavior solutions.
Heterogeneous.vs..Homogenous.Design.of.
Emergent.Specialization
In collective behavior research, approaches to designing emergent special-
ization usually adopt either homogeneous or heterogeneous methods for
designing system components. Homogeneous approaches utilize a single
agent behavior for every agent in a group of agents. Agent behavior may be
encoded as one genotype representation, or in some cases simply defned
by a given set of parameters, which are copied for each agent in the group
(Quinn, Smith, Mayley, & Husbands, 2003). Heterogeneous approaches utilize
different behaviors for each agent in a group of agents. The set of different
behaviors is sometimes encoded as different populations of genotypes, as in
the case of cooperative co-evolutionary genetic algorithms (Parker, 2000).
Alternatively, different agent behaviors may simply be represented as dif-
ferent sets of parameters (Campos et al., 2001).
Designing emergent specialization has been studied via specifying homogene-
ity vs. heterogeneity within both the genotypes and phenotypes of individual
agents as well as entire agent groups. Specialization is often closely associated
with, and sometimes synonymous with, heterogeneity in collective behavior
systems (Balch, 1998; Potter et al., 2001). Heterogeneity can be hardwired
or plastic, and may assume either behavioral (Bryant et al., 2003; Noirot et
al., 1987; Whiteson et al., 2003), or morphological (ORiain, J arvis, Alex-
ander, Buffenstein, & Peeters, 2000; Schultz & Bugajska, 2000; Zhang et
al., 2003) forms. Plastic heterogeneity is when a group adapts its degree of
heterogeneity as a function of environment and task constraints, where as,
hardwired heterogeneity is when the degree of heterogeneity in the group
remains static (Li et al., 2002).
Certain researchers have attempted to outline generalized guidelines as
when to use either homogeneous or heterogeneous design approaches. For
example, Balch (1998) suggested that collective behavior task domains
where all individuals are able to perform the task, such as collective gather-
ing, are particularly suited for homogeneous design. Whilst, task domains
that explicitly require complementary roles, such as RoboCup soccer, are
more suitable for heterogeneous approaches. However, such guidelines, as
with studies of specialization, are usually defned according to the goals and
perspectives of the researcher. Hence, one can readily fnd examples of when
homogeneity and heterogeneity have been used in a manner incongruent to
any given set of design principles or guidelines.
Homogeneous.Approaches
In homogeneous approaches, specialization is typically studied at the group
level since emergent specialization depends upon the local interactions of
cloned behaviors. At the genotype level, the key advantage of a homogeneous
approach is that the search space size is kept minimal since an algorithm need
only optimize a single behavior. At the phenotype level, homogeneous groups
are potentially more adaptive than heterogeneous groups at coping with the
loss of group members. Also, homogenous groups typically have greater
fexibility in coordinating behaviors so as to produce an effective collective
behavior (Stone & Veloso, 1999). The key disadvantage of such approaches
is that system homogeneity, either at the genotype or phenotype level, does
not facilitate specialization, so it is likely that such collective behavior sys-
tems will converge to a non-specialized solution, even if specialization is
advantageous in the given task domain.
Heterogeneous.Approaches
Heterogeneous approaches typically study emergent specialization at either
the local (agent) or global (entire group) level. The key advantage of hetero-
geneity is that it encourages and facilitates emergent specialization, both at
the individual and group level. The key disadvantage of heterogeneous ap-
proaches is that the search space is usually (for complex tasks) prohibitively
large comparative to homogeneous approaches, since many different agent
behaviors need to be optimized or otherwise adapted for task accomplish-
ment.
Collective.Behavior.Tasks.and.Specialization
In the design of collective behavior systems, it remains an open research
question as to which task domains are most appropriately solved using spe-
cialization. However, there is some agreement amongst researchers that if the
task can be naturally decomposed into a set of complementary sub-tasks then
specialization is often benefcial for increasing collective task performance
(Arkin, 1998; Arkin et al., 1999; Balch, 2002a; Balch, 2002b). The following
list enumerates several categories of such collective behavior task domains.
Each of these task domain categories mandates some degree of collective
behavior, where specialization is benefcial for improving task performance.
In subsequent sections, specifc research examples selected from each of
these categories are briefy examined.
Collective.Gathering (Bonabeau et al., 1998; Perez-Uribe, Floreano,
& Keller, 2003).
Collective.Construction (Murciano et al., 1997a, 1997b; Theraulaz &
Bonabeau, 1995).
Collective. Resource. Distribution. and.Allocation (Bonabeau et al.,
1996, 1997; Campos et al., 2001; Theraulaz et al., 1998a, 1998b).
Multi-Agent.Computer.Games (Bryant et al., 2003), (Stanley & Miik-
kulainen, 2002; Stanley et al., 2005b).
RoboCup.Soccer (Luke, Farris, J ackson, & Hendler, 1998; Luke et al.,
1996; Stone et al., 1999.
Predator-Prey.and.Collective.Herding.Behaviors (Blumenthal et al.,
2004b), (Blumenthal & Parker, 2004a, 2004c; Luke et al., 1996; Potter
et al., 2001).
Moving.in.Formation.and.Cooperative.Transportation.Tasks (Kube
& Bonabeau, 1999; Nolf et al., 2003b; Quinn et al., 2003).
Collective.Gathering
Collective gathering is a task domain characterized by the social insect
metaphor. That is, collective gathering tasks seek to emulate the success and
effciency of social insects in gathering resources. Collective gathering tasks
have been studied in the context of both physical multi-robot systems (Kreiger
et al., 2000; Mataric, 1997) and simulated multi-robot systems (Ijspeert,
Martinoli, Billard, & Gambardella, 2001), as well as more abstract artifcial
life simulations (Bongard, 2000; Deneubourg, Theraulaz, & Beckers, 1991;
Perez-Uribe et al., 2003). The collective gathering task domain requires that
a group of agents search for, collect, and transport resources in the environ-
ment from their initial locations to some particular part of the environment.
Such gathering tasks typically require that the group of agents allocate their
labor efforts to particular sub-tasks so as to derive a collective behavior that
maximizes the quantity of resources gathered
6
. Collective gathering tasks
are typically viewed as optimization problems and have been traditionally
studied with mathematical or otherwise analytical methods (Bonabeau et al.,
1996; Gautrais et al., 2002; Theraulaz et al., 1998a).
Learning Behavioral Specialization for Stick Pulling
The research of Li et al. (2004) addressed the important issue of attempting to
specify the concepts of heterogeneity and specialization in a formal defnition,
so as emergent heterogeneity and specialization
7
would be measurable within
the larger context of collective behavior and distributed systems research. In
a case study that compared centralized and distributed learning methods, the
authors qualitatively measured the diversity and specialization of a simulated
multi-robot system given a stick-pulling task that mandated specialized and
cooperative behavior. One research goal was to investigate the impact of
diversity, in the form of heterogeneity in behaviors, upon emergent special-
ization and in turn the impact of specialization on task performance.
In all experiments, the authors presented a learning method that effectively
operated within a multi-robot simulator, where specialization emerged as a
function of task constraints and environmental conditions regardless of whether
local or global reinforcement signals were used. The authors explanation for
this result was that if behavioral diversity (heterogeneity) is benefcial to task
performance, then the learning method facilitates emergent specialization as
a means of taking advantage of this behavioral diversity.
The key criticism of this research is the dependency between emergent
specialization and the learning method used, and consequently the methods
applicability to more generalized optimization tasks. Results supported
a hypothesis that if behavioral diversity in a group was benefcial to task
performance, then specialization was likely to emerge and increase ac-
cordingly with behavioral diversity and task performance. However, these
results largely depended upon the type of learning method, the model of
the task environment, robot controller parameters that defned membership
to a caste, and the task related parameters that the learning method sought
to optimize. Thus, the degree to which emergent specialization depended
upon the underlying adaptation process remains an open question. Also, the
system designer needed to select task environment parameters for the learn-
ing method. This cast doubt upon the possibility of applying the learning
method to more complex and dynamic task environments, where pertinent
task environment parameters that the learning method would require in order
to encourage diversity, specialization, and increased task performance, could
not be identifed a priori.
Furthermore, the number of castes composing a group was determined by the
system designer and not by the adaptive process. Experiments that analyzed
emergent caste formation would be necessary in order to effectively ascer-
tain the relationship between heterogeneity, specialization, and collective
behavior task performance. An adaptive process where a particular number
of castes emerge in response to simulation environment and task constraints
would make such a process applicable to complex task environments where
task challenges dynamically arise.
Collective.Construction
Collective Construction is a task domain characterized by the social insect
metaphor. That is, collective construction tasks seek to emulate the success
and effciency of social insects in gathering resources. Collective construction
tasks have mainly been studied in the context of artifcial life simulations
(Bonabeau, Theraulaz, Arpin, & Sardet, 1994; Murciano et al., 1997a, 1997b;
Theraulaz & Bonabeau, 1995. Collective construction is typically viewed as
an extension of the collective gathering task, in that it requires the agents to
construct a particular structure, with gathered resources, at a home area of
the environment. Specialization is typically required for building complex
structures from many different types of component resources.
Reinforcement Learning for Specialization in Collective Construction
Murciano et al. (1997a) and Murciano et al. (1997b) applied reinforcement
learning (RL) methods to a group of homogeneous agents operating in a
discrete simulation environment. A collective gathering task mandated that
individual agents derive specialized behavior in order to then derive an op-
timal collective behavior.
The authors used a RL method that independently modifed action selection
parameters within the controller of each agent. The RL method used either
global or local RL signals so as to effectuate the learning of specialized
behaviors. Behavioral specialization took the form of an agent learning to
consistently select one action from a set of possible actions. The global RL
signal measured group performance, and the local RL signal measured indi-
vidual performance. The global RL signal was given at the end of a RL trial,
where the signal was equal for all agents in the group. The local RL signal
was given to individual agents, where the signal was calculated in terms of
the agents own successes or failures. Murciano et al. (1997b) conducted
experiments that tested the impact of local versus global RL signals upon
the learning of specialized behaviors in a homogenous group of agents with
no communication. The goal of these experiments was for agents to special-
ize via learning to gather specifc object types so as to construct complex
objects. Thus, when agents interacted an effective collective gathering and
construction behavior emerged. Group task performance was measured as
the number of complex objects assembled in a given RL trial. In the same
experimental setup (Murciano et al., 1997a) conducted experiments that
utilized only global RL signals for the purpose of facilitating emergent spe-
cialization within a homogeneous group of communicating agents. The task
of individual agents and the group was to maximize the number of objects
gathered over the course of a RL trial. The goal of experiments was for agents
to specialize to different behaviors so as communication would facilitate the
collective gathering of an optimal number of objects.
One criticism of the research of Murciano et al., (1997b), and Murciano et
al. (1997a) derives from the use of RL signals in effectuating specialized
behavior. Experimental results indicated that a global RL signal success-
fully motivated emergent specialization, given the assumption that all agents
contribute equally to the task, and the signal was translated so as it could be
meaningfully interpreted by each agent in a homogenous group. This casts
doubt upon the applicability of global RL signals to heterogeneous groups.
Likewise, the applicability of local RL signals was not tested in complex task
domains that provided more realistic simulations of multi-robot systems.
The possibility of applying the RL method to facilitate specialization in
continuous simulation and physical task domains seems unlikely given the
sparse reinforcement limitations of global RL signals and the noisy nature
of local RL signals (Sutton et al., 1998) that inhibit learning. One aim of
the research was to demonstrate that specialization emerges as a function
of task constraints on the environment and agent group, irrespective of the
type of reinforcement signal used. Achieving scalability in the learning of
behavioral specialization is especially prevalent for tasks that require an in-
creasing degree of heterogeneity, and complexity in collective behavior, as a
response to dynamically emerging task challenges. However, the scalability
of the RL method as a mechanism for encouraging behavioral specializa-
tion remains unclear since only two group sizes (10 and 30 agents), and a
discrete environment of one size (54 x 54 grid cells) was tested. Also, the
impact of more dynamic versions of the simulation environment upon the
RL algorithm, were not tested. That is, only one redistribution of objects,
during given RL trials, was tested.
Finally, the RL method assumed that the given task environment could be
abstracted to the form of a multi-objective function which could be opti-
mized. In this case the function was represented as a set of agent affnities
that determined an agents propensity to adopt particular behavioral roles.
This severely limited the applicability of the RL method to more general and
complex task environments.
Collective.Resource.Distribution.and.Allocation
In a series of research endeavors inspired by social insects (Bonabeau et al.,
1996, 1997; Campos et al., 2001; Theraulaz et al., 1998a, 1998b), studied
emergent specialization using response threshold methods in simulations
of homogenous agent groups that were implemented within the context of
mathematical frameworks.
Division of Labor for Dynamic Task Allocation
Theraulaz et al. (1998a) extended a previous formalization for the regulation
of division of labor (Bonabeau et al., 1996) in simulated social insect colonies
so.as to include a reinforcement learning process. A formal variable response.
threshold method was implemented for purpose of facilitating emergent.spe-
cialization in the form of division of labor. The authors highlighted.similarities
between their results and observations made within biological.social systems
where specialist workers were dynamically allocated based.upon sub-task
demand within a collective behavior task (ODonnell, 1998).
Division of Labor for Dynamic Flow Shop Scheduling
Campos et al. (2001) introduced a division of labor method and applied it
as a.method for assigning resources within a dynamic fow shop scheduling
task..The task entailed assigning trucks to paint booths in a factory, where
trucks.moved along an assembly line at a given pace. The color of a truck
was.predetermined by customer order. Three minutes was needed to paint
a truck,.but an additional three minutes was required if the color of a paint
booth was.to be changed for the truck. There was also a cost associated with
paint. changeover for a booth. A division of labor method was applied to
minimize.the number of such changeovers. Such paint ft-and-fnish tasks
are traditional bottleneck problems that can signifcantly reduce production
throughput and thus require optimal solutions (Morley & Ekberg, 1998).
Division of Labor as a Function of Group Size
Gautrais et al., (2002) implemented a variable response threshold method
to.demonstrate that increasing agent group size and demand for tasks gener-
ated.specialized agents. As with previous research (Bonabeau et al., 1996;
Theraulaz et al., 1998a, 1998b), the response threshold method.provided each
agent in a group with an internal threshold for activating a.particular behavior.
Each agents response threshold was infuenced by the.level of demand for a
particular task, and agents allocated themselves so as to.satisfy demand for
these tasks. The authors main conclusion was that their.response threshold
method demonstrated emergent specialization to be.function of group size
in the given resource allocation task, where group sizes.exceeding a criti-
cal threshold value contained specialized agents, and group.sizes below the
critical threshold value contained only unspecialized agents..These fndings
were corroborated by similar fndings in empirical theoretical.biology stud-
ies (Robson et al., 1999).
Division of Labor Methods for Collective Resource Distribution and
Allocation: Comments
Such response threshold methods represent a very simple, yet powerful, self-
regulating feedback system that assigns the appropriate numbers of agents
to different tasks. It is obvious that the study of such biologically inspired
formalizations of specialization are worthy of future research attention given
their applicability to a broad range of optimization tasks including dynamic
scheduling and resource allocation. The methods of Bonabeau et al. (1997),
Campos et al. (2001), Gautrais et al. (2002), and Theraulaz et al. (1998a) were
prevalent in that they eloquently demonstrated how behavioral specialization
emerged as a result of self-regulating task assignment and accomplishment,
for which there exists a large amount of corroborating biological literature and
empirical evidence (Chen, 1937a, 1937b; Deneubourg et al., 1987; ODonnell,
1998; Robson et al., 1999; Theraulaz, Gervet, & Semenoff, 1991).
The main appeal of this set of research examples was their successful modeling
of specialized behavior in the form a set of equations. These equations were
successfully applied as a method for regulating the specialization of agents
to specifc tasks, in order to optimally accomplish a collective behavior task.
However, in many cases the adaptive nature of response threshold regulation
was never tested for more than one group or environment size, and more
than two tasks. Also, the removal of specialized agents to test the adapta-
tion process was limited to two agents. This was an important aspect of the
adaptive nature of response thresholds, since if task allocation becomes too
dynamic, or oscillatory, it is conceivable that the advantages of specialization
could be lost as an agent spends all of its time switching between tasks, and
consequently never dedicates enough time to accomplish a given task.
In each case, a simple set of experiments illustrated the importance and
necessity of utilizing models of biological social behavior as a step towards
understanding such social behavior, and then applying the underlying
techniques, namely response thresholds, as a means of designing problem
solving methods for optimization tasks. The main advantage of division of
labor methods is their eloquence and simplicity of formal specifcation. Also,
such methods yield results that are amenable to a mathematical or formal
analysis. However, such methods are also limited to task domains that can
be completely represented via the mechanics of a mathematical method.
This makes the contributions of such methods limited to optimization tasks
that can be formally represented, or to supporting empirical results evident
in related biological literature.
Multi-Agent.Computer.Games
The application of biologically inspired methods to multi-agent computer
games (Fogel, Hays, & J ohnson, 2004; Laird & vanLent, 2000) has recently
achieved particular success and gained popularity. For example, there has been
particular research interest in the creation of adaptive interactive multi-agent
frst-person shooter games (Cole, Louis, & Miles, 2004; Hong & Cho, 2004;
Stanley et al., 2005b), as well as strategy games (Bryant et al., 2003; Revello
& McCartney, 2002; Yannakakis, Levine, & Hallam, 2004) using artifcial
evolution and learning as design methods for agent behavior. However, the
study of specialized game playing behaviors, in teams of agents, has received
relatively little research attention. Specialization is benefcial since it is often
necessary for teams of agents to formulate collective behavior solutions in
order to effectively challenge a human player, where an increasingly diffcult
level of agent performance is expected as game time progresses.
Legion-I: Neuro-Evolution for Adaptive Teams
Bryant et al. (2003) utilized the Enforced Sub-Populations (ESP) neuro-
evolution method (Gomez, 1997) for the derivation of collective behavior
in a multi-agent strategy game called Legion-I. The research hypothesis was
that a team of homogeneous agents, where agents were capable of adopting
different behavioral roles would be advantageous in terms of task perfor-
mance, comparative to heterogeneous groups, composed of agents with static
complementary behaviors. These experiments highlighted the effectiveness
of the ESP method for deriving a dynamic form of emergent behavioral
specialization motivated by division of labor. Results supported the hypoth-
esis that for the Legion game, a homogenous team, where individuals could
dynamically switch between specialized behaviors was effective. However,
the analysis of emergent specialization was only at a behavioral level, so one
could not readily ascertain the relationship between behavioral specialization
and the evolved genotypes responsible for such behaviors. This would make
an exploration of the mechanisms responsible for emergent specialization
resulting from division of labor problematic. The task environment used a
discrete simulation environment popular in multi-agent strategy games, but
this was not suffciently complex or dynamic in order to adequately test and
support suppositions stating the advantages of behavioral specialization in
homogenous teams. Also, the task performance of homogenous groups was
not compared with heterogeneous groups. Valuable insight into the capabilities
of homogenous versus heterogeneous agent groups for facilitating emergent
specialization, could be gained by a comparison between groups represented
by one neural controller, versus each agent within a group being represented
by a different neural controller.
NERO: Neuro-Evolution of Augmenting Topologies
Stanley et al. (2005b), Stanley, Bryant, Karpov, and Miikkulainen, (2006),
and Stanley, Bryant, and Miikkulainen (2005a) introduced a neuro-evolution
method for the online evolution of neural controllers that operated in the
context of an interactive multi-agent computer game called Neuro-Evolving
Robotic Operatives (NERO). NERO is a frst-person perspective shooter
game, where a human player competes with teams of agents, and agents
compete against each other. The rtNEAT neuro-evolution method was used
for evolving increasing complex agent neural controllers using a process
known as complexifcation. This was an extension of the Neuro-Evolution of
Augmenting Topologies (NEAT) method (Stanley et al., 2002) that operated
using online evolution. The authors demonstrated the effectiveness of the
rtNEAT method for dynamically adapting agent controllers within a team
playing against other agent teams or a human player in real time. Agent
game playing behavior became increasingly sophisticated over successive
generations as a result of changing neural network topological structure as
well as evolving network connection weights. As an extension of the NEAT
method, rtNEAT used online evolution to yield impressive results in terms of
facilitating effectively competitive collective behaviors in the game playing
time of NERO. The NEAT and rtNEAT methods successfully implemented
a speciated representation of the genotype space, and a distance measure
for genotype similarities, that provided a clear method for relating observed
behaviors with a given set of genotypes.
However, the specialized controllers evolved were primarily determined by
a training phase of NERO. Agent teams evolved specializations that were
suitable for a given environment. Given that simulation environments were
the same for both training and a subsequent battle phase, it remains unclear
how suitable evolved teams would be for generalized collective behavior
games. The true potential and benefcial nature of the rtNEAT method for
evolving specialized behaviors in an online evolutionary process, for pur-
pose of increasing team task performance, was not tested in other simulated
multi-robot task domains. In realistic collective behavior tasks where the
environment is dynamic and its structure and layout are not known a priori,
training phases would only be partially effective since controllers trained in
a simulation of the environment would simply be representing a best guess
behavior. Currently, it remains unclear if rtNEAT could be successfully applied
to collective behaviors tasks where there is a signifcant disparity between
a training simulation and a subsequent actual simulation (called the battle
phase in NERO). Such an issue is especially prevalent if online evolution of
controllers is to eventually be applied for accomplishing multi-robot tasks,
with time and energy constraints, in dynamic and complex physical task
environments.
RoboCup.Soccer
A distinct relation to multi-agent game research is RoboCup (Kitano & Asada,
2000). RoboCup is a research feld dedicated to the design and development
of multi-robot systems for the purpose of playing a robotic form of soccer. It
is widely recognized as a specifc test bed for machine learning algorithms,
and engineering challenges (Noda & Stone, 2001). The very nature of the
RoboCup game demands the existence of several types of behavioral spe-
cialization, in the form of different player roles. Such behaviors must be
complementary and able to interact in such a way so as to produce a desired
global behavior. That is, a team strategy that wins the game in a competitive
scenario. Several researchers have focused on machine learning, evolutionary
computation, and neuro-evolution methods that derive task accomplishing
collective behaviors within groups of two or three soccer agents. Although,
specialized behaviors of individual soccer agents was either specifed a priori
or was derived in simplistic game scenarios (Hsu & Gustafson, 2001, 2002;
Luke et al., 1998; Matsubara, Noda, & Hiraki, 1996; Stone et al., 1998,
1998b, 2002; Whiteson et al., 2003. Each of these research examples has
been critiqued elsewhere (Nitschke, 2005).
Pursuit-Evasion
Pursuit-evasion is a collective behavior task that is commonly used within
artifcial life research to test both non-adaptive (typically game theoretic) and
adaptive (typically learning and evolution) methods for agent controller design.
The task requires that multiple pursuer agents derive a collective behavior
for the capture of one or more evading agents (Haynes & Sen, 1996a). The
investigation of emergent specialization remains a relatively unexplored area
of research in the pursuit-evasion domain (Luke et al., 1996), the collective
herding variation (Potter et al., 2001), as well as more traditional predator-
prey systems (Nishimura et al., 1997).
Evolving Pursuit-Evasion Behavior with Hexapod Robots
Blumenthal et al. (2004a, 2004b, 2004c) expanded previous work via com-
bining a punctuated anytime learning (Blumenthal & Parker, 2006; Parker,
2000) method with an evolutionary algorithm within a co-evolution scenario.
Although not the main research focus, this work addressed the issue of using
morphological differences in agents in order to effectuate the derivation of
behavioral specialization, and consequently a collective prey-capture behavior.
The co-evolution scenario operated within a simulated multi-robot system of
fve hexapod robots where the goal was to derive an effective prey-capture
behavior within four predator robots, and a predator-evasion behavior within
one prey robot. This study effectively illustrated the derivation of prey-cap-
ture behavior based upon specialized behaviors that utilized differences in
simulated hexapod robot morphology. Such as, the least maneuverable robot
adopting a passive defensive position, whilst the fastest and most maneuver-
able robots adopted proactive pursuit behaviors.
However, the morphological differences between the robots were simple,
leading one to speculate that a higher degree of complexity in specialized
behavior may have emerged if differences in sensors and controller structure
were included along with a greater disparity in actuator capabilities. Also, the
prey was always initially placed at the center of the simulation environment,
which made it easier for predators to form an effective prey capture behavior,
and infuenced the types of prey-capture behaviors that could emerge. Though
not explicitly stated as a being a goal of this research, a valuable contribution
to this research, would have been a methodological study that described a
mapping or set of principles linking types of sensor and actuator capabilities
to resulting forms of emergent behavioral specialization. Such a study could
potentially form the basis of multi-robot system design methodologies that
use evolution and learning mechanisms that capitalize on morphology in
order to produce desired collective behaviors for solving a given task.
Evolving Herding Behavior in a Multi-Robot System
The research of Potter et al., (2001) investigated the evolution of homogeneous
vs. heterogeneous controllers within a simulated multi-robot system that
was given a collective herding task. A group of Nomad 200s were simulated
within the TeamBots simulator (Balch, 1998). The research hypothesis was
that as task diffculty increased, heterogeneity and specialization become
essential for successful task accomplishment. Heterogeneity was defned
as the number of different behaviors one robot could select from, as well
as the number of behaviors in the group. This hypothesis was tested with
experiments that introduced a predator into the environment. The goal was
to encourage the emergence of specialized defensive behaviors in addition to
herding behaviors. Experiments effectively illustrated that emergent behav-
ioral specialization, for the beneft of collective behavior task performance,
could be facilitated in a heterogeneous team of agents. Furthermore, results
supported a hypothesis that constructing a collective behavior task such that
multiple behaviors are required, increases the need for heterogeneity, and in
turn specialization. However, the inducement of emergent specialization via
increasing the number of behaviors required, and not simply task complexity,
was only investigated within a single case study.
The key criticism lies in the comparison of homogenous and heterogeneous
groups for deriving collective herding behaviors. Particularly, it is unclear
why the authors opted to use only two genotype populations to represent a
group of three shepherds in the heterogeneous design approach. The impact of
homogeneity and heterogeneity on emergent specialization was not validated
with larger groups of shepherds. Also, only one increment in the complexity
of the task environment was tested. That is, the addition of the predator to the
collective herding task. Complete validation of the authors hypothesis that
specialization emerges not as a consequence of task complexity, but rather as
a result of the number of behaviors required to solve the task, would require
several comparative case studies. Such studies would need to test tasks of
varying degrees of diffculty versus tasks that require numerous complementary
and potentially specialized behaviors. Such a comprehensive study would
yield a valuable contribution to ones understanding of the relation between
heterogeneous and homogenous design approaches, task performance, task
complexity, and emergent specialization.
Moving.in.Formation.and.Cooperative.Transportation.
Tasks
Certain collective behavior research endeavors, mainly in the felds of artifcial
life and multi-robot systems, have aimed to model and reproduce various forms
of social phenomena that are observable in biological systems (Reynolds,
1987; Zaera, Cliff, & Bruten, 1996). Coordinated movement and cooperative
transport is sometimes studied within the context of a gathering task, and
has been studied separately in both physical and simulated environments.
Cooperative transport is inspired by biological prey retrieval models, which
present many examples of the value of specialization, such as the pushing vs.
pulling behaviors exhibited in stigmatic coordination that allows several ants
to transport a large prey (Kube et al., 1999). Such inspiration was used by the
research of Dorigo et al. (2004) and Nolf, Baldassarre, and Parisi (2003a),
which described the evolution of coordinated motion, and self-assembly in
a simulated multi-robot system for the purpose of cooperatively transporting
objects. Similarly, the research of (Nolf et al., 2003a) described the evolution
of particular group formations in a simulated multi-robot system, which al-
lowed effcient forms of coordinated group movement across an environment
towards a light or sound source. The research of Baldassarre, Nolf, and Parisi
(2003), Dorigo et al. (2004), and Nolf et al. (2003a) has been reviewed in
related work (Nitschke, 2005), and is thus not described here.
Future.Directions
Consequent of the literature reviewed, we deem the most viable future re-
search direction to be the development of structured and principled emergent
behavior design methodologies. From a broad range of methods that utilize
emergent specialization for solving collective behavior tasks, a lack of a
unifying set of design principles (methodologies) that link the workings of
each of these methods, was highlighted. Such design methodologies would
provide defnitions and measures of specialization, and allow researchers to
construct collective behavior systems that facilitate desired forms of emergent
specialization that solve given tasks. If emergent specialization is to be utilized
as a problem solver in systems that are designed using biologically inspired
principles such as evolution and learning, then the concept of specialization
must be defned, so as it can be identifed and used in a problem solving
process. In order to validate design methodologies that identify, measure,
and harness emergent specialization as a problem solving tool in artifcial
complex adaptive systems, several considerations must be made.
1. Given the disparate and disjoint nature of biologically inspired and col-
lective behavior research, validation of emergent specialization design
methodologies would be experimental, and not necessarily constructed
from a set of mathematical or otherwise theoretical suppositions that
are proved.
2. Such methodologies would need to encapsulate the various types of
specialization that beneft particular types of collective behavior tasks.
These types would be identifed through extensive experimentation.
3. Such methodologies would need to use specialization that can be iden-
tifed and categorized, either dynamically by the design method, or a
priori by a human designer. Importantly, dynamic identifcation of the
type and degree of specialization required for a given task by a method
would greatly increase the applicability of the method. That is, such
a method would by applicable to complex task environments where
specifc task challenges dynamically arise in the environment and the
exact nature of tasks cannot be described ahead of time.
Hence, if emergent specialized behavior is to be used as a means of deriv-
ing solutions to complex and dynamic task challenges in both simulated and
physical collective behavior systems
8
then future research is obliged to look
towards addressing the considerations delineated herein.
Conclusion
In drawing conclusions for this chapter, it is important to note that the chapters
goal was not to present an exhaustive list of research relating to emergent
specialization, but rather to identify and present a set of pertinent research
examples that use biologically inspired design approaches for the purpose of
facilitating emergent behavioral specialization. Such research examples were
selected based upon results that indicated emergent behavioral specialization
as being benefcial for solving collective behavior tasks.
The binding theme of the chapter argued, that the majority of collective
behavior research is currently analyzed and evaluated from empirical data
gathered and emergent behavioral specialization observed, without analytical
methods for identifying the means and causes of emergent specialization. An
obvious reason for this is that the use of biologically inspired concepts such
as evolution, self-organization, and learning as design methods is still in a
phase of research infancy. Consequently, emergent specialization derived
using such biologically inspired design concepts is currently constrained
to simple forms. Given this general evaluation of prevalent literature, we
identifed several unresolved issues that inhibit the development of biologi-
cally inspired design methodologies that synthesize emergent specialization
in solving collective behavior tasks.
1. It was evident that many researchers deem the simulation of collective
behavior systems to be an effective approach for investigating emergent
behavioral specialization, given that simulations provide a convenient
means for studying the conditions under which specialization emerges.
For example, the effects of parametric changes can be observed in a
relatively short space of time. However, with notable exceptions, such
as SwarmBots (Dorigo et al., 2004), the identifcation and transference
of mechanisms motivating emergent specialization observed in simula-
tion to counter-part algorithms operating in physical collective behavior
systems such as multi-robot systems, is not yet plausible. In the case of
SwarmBots (Dorigo et al., 2004), a simple task environment made the
transference to a physical environment possible, and emergent special-
ization was not necessarily a problem solver for dynamic challenges in
the environment, but rather a solution to a given task that was emergent
but not necessarily desired.
2. In the pertinent research examples reviewed, the complete potential
of biologically inspired design, and the advantages of emergent spe-
cialization were not always effectively exploited. For example, many
collective behavior systems, with notable exceptions such as division
of labor methods applied to optimization tasks (Bonabeau et al., 1997),
were simply attempting to synthesize emergent specialization, or to
demonstrate the veracity of concepts such as self-organization, learn-
ing, and evolution for deriving novel agent behaviors. Such concepts
were rarely applied to methods that derived emergent specialization as
a means of increasing task performance or accomplishing unforeseen
challenges in collective behavior tasks.
3. There is currently no standardized benchmark or research test-bed for
testing, interpreting, evaluating, and classifying emergent specialized
behavior. RoboCup was included as an honorable mention in the chap-
ter, given that it provides an effective platform for testing and evaluat-
ing various forms of collective and individual behavior, emergent or
otherwise, implemented either within an agent simulator or a physical
multi-robot system. That is, collective behavior is simply evaluated
within a competitive game scenario, so collective behavior performance
is determined according to the evaluation criteria of the game. Another
exception is collective gathering and dynamic scheduling in distributed
systems, which can be represented as optimization tasks. In this case,
standardized benchmarks exist in the form of performance results yielded
by classical adaptive approaches. This makes the results of biologically
inspired and classical methods to such tasks comparable. However, with
exceptions such as Bonabeau et al. (1997) and Theraulaz and Bonabeau
(1995) many optimization tasks do not beneft from the use of emergent
behavioral specialization. Thus, the testing, interpretation, and evalua-
tion of emergent specialized behavior within the context of collective
behavior systems, is currently conducted according to the performance
benchmarks of the researchers own experimental simulation platform.
This means that the experimental results can only be compared within
the context of their own simulation environment. The development of
emergent specialization design methodologies that could be equally ap-
plied to physical collective behavior systems would remove this critical
constraint.
Given these open research issues, one may conclude that if the notion of
emergent specialization as a problem solver for collective behavior tasks is
to gain any maturity and credibility, then collective behavior systems must
be built upon proven emergent specialization design methodologies. Ideally,
such methodologies must be proven for convergence to desired forms of
collective behavior (achieved as a consequence of emergent specialization),
scalable and transferable to a counterpart situated and embodied collective
behavior task environments.
References
Abdel-Rahman, H. (2001). When do cities specialize in production. Reg Sci
Urban Econ, 26(1), 1-22.
Arkin, R. (1998). Behavior based robotics. Cambridge, MA: MIT Press.
Arkin, R., & Balch, T. (1999). Behavior-based formation control for multi-
robot teams. IEEE Transactions on Robotics and Automation, 14(6),
926-939.
Axelrod, R. (1984). The evolution of cooperation. New York: Basic
Books.
Baev, K. (1997). Biological neural networks. Berlin, Germany: Birkuser.
Balch, T. (2002b). Taxonomies of multi-robot task and reward. In T. Balch
& E. Parker (Eds.), Robot teams: From diversity to polymorphism (pp.
23-35). Natick: A K Peters.
Balch, T. (2002a). Measuring robot group diversity. In T. Balch & E. Parker
(Eds.), Robot teams: From diversity to polymorphism (pp. 93-135).
Natick: A K Peters.
Balch, T. (1998). Behavioral diversity in learning robot teams. PhD thesis.
Atlanta: College of Computing, Georgia Institute of Technology.
Baldassarre, G., Nolf, S., & Parisi, D. (2003). Evolving mobile robots able
to display collective behavior. ArtifcialLife, 9(3), 255-267.
Beni, G. (2004). From swarm intelligence to swarm robotics. Proceedings
of the 1
st
International Workshop on Swarm Robotics (p. 1-9). Santa
Monica, CA: Springer.
Blumenthal, J ., & Parker, G. (2006). Benchmarking punctuated anytime
learning for evolving a multi-agent teams binary controllers. World
AutomationCongress2006.Budapest, Hungary: IEEE Press.
Blumenthal, J ., & Parker, G. (2004a). Co-evolving team capture strategies
for dissimilar robots. AAAIArtifcialMulti-AgentLearningSymposium
(pp. 15-23). Arlington, VA: AAAI Press.
Blumenthal, J ., & Parker, G. (2004b). Competing sample sizes for the co-
evolution of heterogeneous agents. Proceedingsofthe2004IEEE/RSJ
International Conference on Intelligent Robots and Systems (pp. 1438-
1443). Sendai, J apan: IEEE Press.
Blumenthal, J., & Parker, G. (2004c). Punctuated anytime learning for evolving
multi-agent capture strategies. Proceedings of the Congress on Evolu-
tionary Computation (pp. 1820-1827). Portland, OR: IEEE Press.
Bonabeau, E., & Theraulaz, G. (1999). Role and variability of response
thresholds in the regulation of division of labor in insect societies. In
J . Deneubourg & J . Pasteels (Eds.), Information processing in social
insects (p. 141-163). Basel, Switzerland: Springer Verlag.
Bonabeau, E., Dorigo, M., & Theraulaz, G. (1998). Swarm intelligence:
Fromnaturaltoartifcialsystems.Oxford, England: Oxford University
Press.
Bonabeau, E., Sobkowski, A., Theraulaz, G., & Deneubourg, J . (1997). Adap-
tive task allocation inspired by a model of division of labor in social
insects. In D. Lundh, B. Olsson, & A. Narayanan (Eds.), Bio-computing
and emergent computation (pp. 36-45). Singapore: World Scientifc.
Bonabeau, E., Theraulaz, G., & Deneubourg, J . (1996). Quantitative study
of the fxed threshold model for the regulation of division of labor in
insect societies. Proceedings of the Royal Society of London B, 263(1),
1565-1569.
Bonabeau, E., Theraulaz, G., Arpin, E., & Sardet, E. (1994). The building
behavior of lattice swarms. ArtifcialLifeIV:Proceedingsofthe4
th
In-
ternational Workshop on the Synthesis and Simulation of Living Systems
(pp. 307-312). Cambridge, MA: MIT Press.
Bongard, J . (2000). The legion system: A novel approach to evolving het-
erogeneity for collective problem solving. ProceedingsofEurogp-2000
(pp. 16-28). Edinburgh, UK: Springer-Verlag.
Brooks, R. (1990). Elephants dont play chess. Robotics and Autonomous
Systems, 6(1), 3-15.
Bryant, B., & Miikkulainen, R. (2003). Neuro-evolution for adaptive teams.
Proceedingsofthe2003CongressonEvolutionaryComputation(pp.
2194-2201). Canberra, Australia: IEEE Press.
Calderone, N., & Page, R. (1988). Genotypic variability in age polyethism
and task specialization in the honeybee. Apis mellifera. Behav. Ecol.
Sociobiol, 22(1), 17-25.
Campos, C., Theraulaz, G., Bonabeau, E., & Deneubourg, J . (2001). Dynamic
scheduling and division of labor in social insects. Adaptive Behavior,
8(2), 83-94.
Chen, C. (1937a). The leaders and followers among the ants in nest-building.
Physiol. Zool., 10(1), 437-455.
Chen, C. (1937b). Social modifcation of the activity of ants in nest-building.
Physiol. Zool., 10(1), 420-436.
Cole, N., Louis, S., & Miles, C. (2004). Using a genetic algorithm to tune
frst-person shooter bots. Proceedingsofthe2004CongressonEvolu-
tionary Computation (Vol. 1, pp. 139-145). Piscataway: IEEE Press.
Deneubourg, J ., Goss, S., Pasteels, J ., Fresneau, D., & Lachaud, J . (1987).
Self-organization mechanisms in ant societies (ii): Learning in forag-
ing and division of labor. In J . Pasteels & J . Deneubourg (Eds.), From
individual to collective behavior in social insects (pp. 177-196). Basel,
Switzerland: Birkhauser.
Deneubourg, J ., Theraulaz, G., & Beckers, R. (1991). Swarm-made architec-
tures. ProceedingsoftheEuropeanConferenceonArtifcialLife(pp.
123-133). Amsterdam, Holland: Elsevier Academic Publishers.
Dorigo, M., Tuci, E., Gross, R., Trianni, V., Labella, H., Nouyan, S., et al.
(2004). The swarm-bots project. Proceedings of the Swarm Robotics:
SAB2004 International Workshop (Vol. 3342, pp. 31-32). Santa Monica,
CA: Springer Verlag.
Edmonds, J. (1965). Path, trees, and fowers. Canadian J. Math., 17(1),
449-467.
Fogel, D., Hays, T., & J ohnson, D. (2004). A platform for evolving characters
in competitive games. Proceedingsof2004congressonevolutionary
computation (pp. 1420-1426). Piscataway, USA: IEEE Press.
Funes, P., Orme, B., & Bonabeau, E. (2003). Evolving emergent group
behaviors for simple humans agents. Proceedingsofthe7
th
European
ConferenceonArtifcialLife(pp. 76-89). Berlin, Germany: Springer-
Verlag.
Gautrais, J ., Theraulaz, G., Deneubourg, J ., & Anderson, C. (2002). Emergent
polyethism as a consequence of increased colony size in insect societies.
Journal of Theoretical Biology, 215(1), 363-373.
Gomez, F. (1997). Robust non-linear control through neuro-evolution. PhD
thesis. Austin, TX: Department of Computer Science, University of
Austin, Texas.
Grefenstette, J . (1995). Credit assignment in rule discovery systems. Machine
Learning, 3(3), 225-246.
Hawthorne, D. (2001). Genetic linkage of ecological specialization and re-
productive isolation in pea aphids. Nature, 412(1), 904-907.
Haynes, T., & Sen, S. (1996a). Co-adaptation in a team. International Journal
of Computational Intelligence and Organizations, 1(4), 1-20.
Haynes, T., & Sen, S. (1996b). Evolving behavioral strategies in predators
and prey. In G. Weiss & S. Sen (Eds.), Adaptation and learning in
multi-agent systems: Lecture notes in computer science (pp. 113-126).
Berlin, Germany: Springer-Verlag.
Hong, J ., & Cho, S. (2004). Evolution of emergent behaviors for shooting
game characters in RoboCode. Proceedingsofthe2004Congresson
Evolutionary Computation (Vol. 1, pp. 634-638). Piscataway, USA:
IEEE Press.
Hsu, W., & Gustafson, S. (2001). Layered learning in genetic programming
for a cooperative robot soccer problem. Proceedings of the 4
th
Euro-
pean Conference on Genetic Programming (pp. 291-301). Como, Italy:
Springer-Verlag.
Hsu,W., & Gustafson, S. (2002). Genetic programming and multi-agent
layered learning by reinforcements. Proceedings of the Genetic and
Evolutionary Computation Conference (pp. 764-771). New York: Mor-
gan Kaufmann.
Ijspeert, A., Martinoli, A., Billard, A., & Gambardella, L. (2001). Col-
laboration through the exploitation of local interactions in autonomous
collective robotics: The stick pulling experiment. Autonomous Robots,
11(2), 149-171.
Kitano, H., & Asada, M. (2000). The RoboCup humanoid challenge as the
millennium challenge for advanced robotics. Advanced Robotics, 13(1),
723-736.
Komann, M., & Fey, D. (2007). Realizing emergent image pre-processing tasks
in cellular-automaton-alike massively parallel hardware. International
Journal on Parallel, Emergent and Distributed Systems, 22(1), 79-89.
Kreiger, M., & Billeter, J . (2000). The call of duty: Self-organized task al-
location in a population of up to twelve mobile robots. Robotics and
Autonomous Systems, 30, 65-84.
Kube, C., & Bonabeau, E. (1999). Cooperative transport by ants and robots.
Robotics and Autonomous Systems: Special Issue on Biomimetic Robots,
1(1), 20-29.
Laird, J ., & vanLent, M. (2000). Human-level AI killer application: Interac-
tive computer games. Proceedingsofthe17
th
National Conference on
Artifcial Intelligence and the 12
th
Annual Conference on Innovative
Applications of Artifcial Intelligence (pp. 1171-1178). Menlo Park,
USA: AAAI Press.
Li, L., Martinoli, A., & Mostafa, Y. (2002). Emergent specialization in swarm
systems. In H. Yin, N. Allinson, R. Freeman, J . Keane, & S. Hubbard
(Eds.), Lecturenotesincomputerscience:Vol.2412.Intelligentdata
engineering and automated learning (pp. 261-266). Berlin, Germany:
Springer-Verlag.
Li, L., Martinoli, A., & Yaser, A. (2004). Learning and measuring specialization
in collaborative swarm systems. Adaptive Behavior, 12(3), 199-212.
Lipson, H., & Pollack, J . (2000). Automatic design and manufacture of robotic
life forms. Nature, 406(1), 974-978.
Luke, S., & Spector, L. (1996). Evolving teamwork and coordination with
genetic programming. Proceedings of the 1996 International Conference
on Genetic Programming (pp. 150-156). Stanford, USA: MIT Press.
Luke, S., Farris, C. H. J ., J ackson, G., & Hendler, J . (1998). Co-evolving
soccer SoftBot team coordination with genetic programming. In H.
Kitano (Ed.), Robocup-97:RobotSoccerWorldCupI(pp. 398-411).
Berlin, Germany: Springer-Verlag.
Martinoli, A., Zhang, Y., Prakash, P., Antonsson, E., & Olney, R. (2002).
Towards evolutionary design of intelligent transportation systems. The
11
th
International Symposium on New Technologies for Advanced Driver
Assistance Systems (pp. 283-290). Siena, Italy: ATA Press.
Mataric, M. (1997). Behavior-based control: Examples from navigation,
learning, and group behavior. Journal of Experimental and Theoretical
ArtifcialIntelligence, 9(1), 62-78.
Matsubara, H., Noda, I., & Hiraki, K. (1996). Learning of cooperative actions
in multi-agent systems: A case study of pass and play in soccer. Adapta-
tion, Co-Evolution, and Learning in Multi-Agent Systems: Papers from
the 1996 AAAI Spring Symposium (pp. 63-67). Boston: AAAI Press.
Morley, R., & Ekberg, G. (1998). Cases in chaos: Complexity-based ap-
proaches to manufacturing. In Embracing complexity: A colloquium on
the application of complex adaptive systems to business. (p. 97-702).
Cambridge, USA: The Ernst and Young Center for Business for Busi-
ness Innovation.
Mller, C., & Sick, B. (2006). Emergence in organic computing systems:
Discussion of a controversial concept. Proceedingsofthe3
rd
Interna-
tionalConferenceonAutonomicandTrustedComputing,Volume4158
Of Lecture Notes In Computer Science (pp. 1-16). Berlin, Germany:
Springer.
Murciano, A., & Millan, J . (1997). Learning signaling behaviors and special-
ization in cooperative agents. Adaptive Behavior, 5(1), 5-28.
Murciano, A., Millan, J., & Zamora, J. (1997). Specialization in multi-agent
systems through learning. Biological Cybernetics, 76(1), 375-382.
Ng, Y., & Yang, X. (1997). Specialization, information, and growth: A se-
quential equilibrium analysis. Rev Dev Econ, 1(1), 257-274.
Nishimura, S., & Takashi, I. (1997). Emergence of collective strategies in a
predator-prey game model. ArtifcialLife, 3(1), 243-260.
Nitschke, G. (2005). Emergence of cooperation: State of the art. Artifcial
Life, 11(3), 367-396.
Noda, I., & Stone, P. (2001). The RoboCup soccer server and CMUnited
clients: Implemented infrastructure for MAS research. Autonomous
Agents and Multi-Agent Systems, 7(1), 101-120.
Noirot, C., & Pasteels, J . (1987). Ontogenetic development and the evolution
of the worker caste in termites. Experientia., 43(1), 851-860.
Nolf, S., & Floreano, D. (2000). Evolutionary robotics: The biology, intel-
ligence, and technology of self-organizing machines. Cambridge, MA:
MIT Press.
Nolf, S., & Parisi, D. (1997). Learning to adapt to changing environments
in evolving neural networks. Adaptive Behavior, 1(5), 75-98.
Nolf, S., Baldassarre, G., & Parisi, D. (2003a). Evolution of collective be-
havior in a team of physically linked robots. In G. Guillot & J . Meyer
(Eds.), Applications of evolutionary computing (pp. 581-592). Heidel-
berg, Germany: Springer Verlag.
Nolf, S., Deneubourg, J., Floreano, D., Gambardella, L., Mondada, F., &
Dorigo, M. (2003b). Swarm-bots: Swarm of mobile robots able to self-
assemble and self-organize. Ecrim News, 53(1), 25-26.
ODonnell, S. (1998). Effects of experimental forager removals on division
of labour in the primitively eusocial wasp polistes instabilis. Behaviour,
135(2), 173-193.
ORiain, M., J arvis, J ., Alexander, R., Buffenstein, R., & Peeters, C. (2000).
Morphological castes in a vertebrate. Proceedings of the National
Academy of Sciences of the United States of America, 97(24), 13194-
13197.
Parker, G. (2000). Co-evolving model parameters for anytime learning in evo-
lutionary robotics. Robotics and Autonomous Systems, 33(1), 13-30.
Parker, G., & Nathan, P. (2006). Evolving sensor morphology on a legged
robot in niche environments. WorldAutomationCongress2006.Buda-
pest, Hungary: IEEE Press.
Perez-Uribe, A., Floreano, D., & Keller, L. (2003). Effects of group composi-
tion and level of selection in the evolution of cooperation in artifcial ants.
AdvancesofArtifcialLife:Proceedingsofthe7
th
European Conference
on ArtifcialLife(pp. 128-137). Dortmund, Germany: Springer.
Pfeifer, R., Iida, F., & Gomez, G. (2006). Designing intelligent robots: On
the implications of embodiment. Journal of Robotics Society of Japan,
24(7), 9-16.
Potter, M., & DeJ ong, K. (2000). Cooperative co-evolution: An architecture
for evolving co-adapted subcomponents. Evolutionary Computation,
8(1), 1-29.
Potter, M., Meeden, L., & Schultz, A. (2001). Heterogeneity in the coevolved
behaviors of mobile robots: The emergence of specialists. Proceedings
of the International Joint Conference on Artifcial Intelligence (pp.
1337-1343). Seattle, WA: AAAI Press.
Quinn, M., Smith, L., Mayley, G., & Husbands, P. (2003). Evolving controllers
for a homogeneous system of physical robots: Structured cooperation
with minimal sensors. Philosophical Transactions of the Royal Society
of London, Series A: Mathematical, Physical, and Engineering Sciences,
361(1), 2321-2344.
Resnick, M. (1997). Turtles,termites,andtraffcjams:Explorationsinmas-
sively parallel micro worlds. Cambridge, MA: MIT Press.
Revello, T., & McCartney, R. (2002). Generating war game strategies using
a genetic algorithm. Proceedingsofthe2002CongressonEvolutionary
Computation (pp. 1086-1091). Piscataway, USA: IEEE Press.
Reynolds, C. (1987). Flocks, herds, and schools: A distributed behavioral
model. Computer Graphics, 21(4), 25-36.
Robson, S., & Traniello, J . (1999). Key individuals and the organization of
labor in ants. In J . Deneubourg & J . Pasteels (Eds.), Information pro-
cessing in social insects (pp. 239-259). Basel, Switzerland: Springer
Verlag.
Schultz, A., & Bugajska, M. (2000). Co-evolution of form and function in
the design of autonomous agents: Micro air vehicles project. Proceed-
ings of the Workshop on Evolution of Sensors in Nature, Hardware, and
Simulation (pp. 154-166). Chicago, IL: AAAI Press.
Seligmann, H. (1999). Resource partition history and evolutionary specializa-
tion of subunits in complex systems. Biosystems, 51(1), 31-39.
Sims, K. (2004). Evolving 3d morphology and behavior by competition.
ArtifcialLifeIV:Proceedingsofthe4
th
International Workshop on the
Synthesis and Simulation of Living Systems (pp. 28-39). Cambridge,
MA: MIT Press.
Smith, A. (1904). An inquiry into the nature and causes of the wealth of na-
tions (5
th
ed.), (E. Cannan, Ed.). London, United Kingdom: Methuen
and Co., Ltd. First published: 1776.
Solow, D., & Szmerekovsky, J . (2004). Mathematical models for explain-
ing the emergence of specialization in performing tasks. Complexity,
10(1), 37-48.
Stanley, K., & Miikkulainen, R. (2002). Evolving neural networks through
augmenting topologies. Evolutionary Computation, 10(2), 99-127.
Stanley, K., Bryant, B., & Miikkulainen, R. (2005a). Evolving neural net-
work agents in the NERO video game. ProceedingsoftheIEEE2005
Symposium on Computational Intelligence And Games (pp. 182-189).
Piscataway, USA: IEEE Press.
Stanley, K., Bryant, B., & Miikkulainen, R. (2005b). Real-time neuro-evolution
in the NERO video game. Evolutionary Computation, 9(6), 653-668.
Stanley, K., Bryant, B., Karpov, I., & Miikkulainen, R. (2006). Real-time
evolution of neural networks in the NERO video game. Proceedings of
the 21
st
NationalConferenceonArtifcialIntelligence(pp. 1671-1674).
Boston: AAAI Press.
Stone, P., & Veloso, M. (1998a). A layered approach to learning client be-
haviors in the RoboCup soccer server. AppliedArtifcialIntelligence,
12(1), 165-188.
Stone, P., & Veloso, M. (1998b). Using decision tree confdence factors for
multi-agent control. Proceedingsofthe2
nd
International Conference on
Autonomous Agents (pp. 110-116). Minneapolis, MN: ACM Press.
Stone, P., & Veloso, M. (1999). Task decomposition, dynamic role assignment,
and low bandwidth communication for real time strategic teamwork.
ArtifcialIntelligence, 110(2), 241-273.
Stone, P., & Veloso, M. (2002). Towards collaborative and adversarial learn-
ing: A case study in robotic soccer. Evolution and learning in multi-agent
systems, 48(1), 83-104.
Sutton, R., & Barto, A. (1998). An introduction to reinforcement learning.
Cambridge, MA: J ohn Wiley and Sons.
Tarapore, D., Floreano, D., & Keller, L. (2006). Infuence of the level of
polyandry and genetic architecture on division of labor. The10
th
Inter-
national Conference on the Simulation and Synthesis of Living Systems
(ALIFE X) (pp. 358-364). Cambridge, USA: MIT Press.
Theraulaz, G., & Bonabeau, E. (1995). Coordination in distributed building.
Science, 269(1), 686-688.
Theraulaz, G., Bonabeau, E., & Deneubourg, J . (1998a). Fixed response
thresholds and the regulation of division of labor in insect societies.
Bulletin of Mathematical Biology, 60(1), 753-807.
Theraulaz, G., Bonabeau, E., & Deneubourg, J . (1998b). Response threshold
reinforcement and division of labor in insect societies. Proceedings of
the Royal Society of London B, 265(1), 327-332.
Theraulaz, G., Gervet, J ., & Semenoff, S. (1991). Social regulation of for-
aging activities in polistes dominulus christ: A systemic approach to
behavioural organization. Behaviour, 116(1), 292-320.
Theraulaz, G., Goss, S., Gervet, J ., & Deneubourg, J . (1991). Task differen-
tiation in polistes wasp colonies: A model for self-organizing groups of
robots. The 1
st
International Conference on the Simulation of Adaptive
Behavior (pp. 346-355). Cambridge, USA: MIT Press.
Thompson, A., Harvey, I., & Husbands, P. (1996). Unconstrained evolution
and hard consequences. Towards Evolvable Hardware: The Evolution-
ary Engineering Approach (Vol. 1062 LNCS, pp. 135-165). Berlin,
Germany: Springer-Verlag.
Trumler, W., Klaus, R., & Ungerer, T. (2006). Self-confguration via coop-
erative social behavior. Proceedingsofthe3
rd
International Conference
on Autonomic and Trusted Computing (pp. 90-99). Wuhan, China:
Springer.
Waibel, M., Floreano, D., Magnenat, S., & Keller, L. (2006). Division of labor
and colony effciency in social insects: Effects of interactions between
genetic architecture, colony kin structure, and rate of perturbations.
Proceedings of the Royal Society B, 273(1), 1815-1823.
Watson, R., Ficici, S., & Pollack, J . (1999a). Embodied evolution: A response
to challenges in evolutionary robotics. In J . Wyatt & J . Demiris (Eds.),
The 8
th
European Workshop on Learning Robots (pp. 14-22). Lausanne,
Switzerland: Springer Verlag.
Watson, R., Ficici, S., & Pollack, J . (1999b). Embodied evolution: Embody-
ing an evolutionary algorithm in a population of robots. 1999 Congress
on Evolutionary Computation (pp. 335-342). Washington, DC, USA:
IEEE Press.
Watson, R., Ficici, S., & Pollack, J . (2002). Embodied evolution: Distribut-
ing an evolutionary algorithm in a population of robots. Robotics and
Autonomous Systems, 39(1), 1-18.
Weis, T., Parzyjegla, H., J aeger, M., & Mhl, G. (2006). Self-organizing and
self-stabilizing role assignment in sensor/actuator networks. Proceedings
ofthe8
th
International Symposium on Distributed Objects and Applica-
tions. (pp. 1807-1824), Montpellier, France: Springer.
Wenseleers, T., Ratnieks, F., & Billen, J. (2003). Caste fate confict in swarm-
founding social hymenoptera: An inclusive ftness analysis. Evolutionary
Biology, 16(1), 647-658.
Whiteson, S., Kohl, N., Miikkulainen, R., & Stone, P. (2003). Evolving
keep-away soccer players through task decomposition. Proceeding of
the Genetic and Evolutionary Computation Conference (pp. 356-368).
Chicago, IL: AAAI Press.
Wu, J., Di, Z., & Yang, Z. (2003). Division of labor as the result of phase
transition. Physica A, 7(1), 323-663.
Yannakakis, G., Levine, J ., & Hallam, J . (2004). An evolutionary approach
for interactive computer games. Proceedings of the 2004 Congress
on Evolutionary Computation (pp. 986-993). Piscataway, USA: IEEE
Press.
Zaera, N., Cliff, D., & Bruten, J. (1996). (Not) Evolving collective behav-
iors in synthetic fsh (tech. rep.). Bristol, England: Hewlett-Packard
Laboratories.
Zhang, Y., Martinoli, A., & Antonsson, E. (2003). Evolutionary design of a
collective sensory system. The2003AAAISpringSymposiumon Com-
putational Synthesis (pp. 283-290). Stanford, USA: AAAI Press.
Endnotes
1
Examples of complex adaptive systems include social insect colonies,
biological neural networks, traffc jams, economies of a nation, as well
as industrial infrastructures such as energy and telecommunications
networks (Resnick, 1997). We deem complex adaptive systems to be a
subset of complex systems where autonomous software (simulated) or
physically embodied (robots) agents operate in order to solve a given
task.
2
The terms task, activity, role, and caste are defned as follows. Task:
what has to be done; Activity: what is being done; Role: the task as-
signed to an individual within a set of responsibilities given to a group
of individuals; Caste: a group of individuals specialized in the same
role (Kreiger et al., 2000).
3
The terms collective behavior system and artifcial complex adaptive
system are used interchangeably throughout the chapter. Both refer to
distributed systems where specialization emerges as a property of a
collective behavior dynamics.
4
We distinguish methodologies from methods. We assume the latter to be
the actual algorithm, which is implemented for the purpose of solving
a specifc task. Where as, we assume the former to be a set of design
principles for designing methods.
5
Various defnitions for numerous types of specialization have been
proposed across a broad range of disciplines. In The Wealth of Nations,
(Smith, 1904) Adam Smith described economic specialization in terms
of division of labor. Specifcally stating that in industrialism, division
of labor represents a qualitative increase in productivity, and regarded
its emergence as the result of a dynamic engine of economic progress.
Smith viewed specialization by workers as leading to greater skill and
greater productivity for given tasks, which could not be achieved by
non-specialized workers attempting to accomplish those same tasks.
6
The allocation of agent labor within a group of agents is analogous
to resource allocation which derives from economic and game theory
studies (Axelrod, 1984). Such studies attempt to derive methods that
effciently allocate a limited amount of resources so as to accomplish a
given task with the highest degree of performance possible.
7
Heterogeneity, and hence behavioral diversity, was defned as the number
of castes in the group, and specialization was the part of diversity that
was required to increase task performance.
8
Such a case has been envisioned for swarm robotic systems (Beni,
2004).
Section IV
Social Science
Perspectives
Emergence n Agent-Based Computatonal Socal Scence
Chapter.IX
Emergence.in.Agent-Based.
Computational.Social.
Science:
Conceptual,.Formal,.and.
Diagrammatic.Analysis
J ean Lous Dessalles, ENST, Pars, France
J acques Ferber, LIRMM, CNRS, & Unversty of Montpeller, France
Dens Phan, GEMAS, CNRS, and Unversty of Pars IV Sorbonne,
& CREM, CNRS, and Unversty of Rennes I, France
Abstract
Thischapterprovidesacriticalsurveyofemergencedefnitionsbothfrom
a conceptual and formal standpoint. The notions of downward/backward
causation and weak/strong emergence are specially discussed for applica-
tion to complex social system with cognitive agents. Particular attention is
devotedtotheformaldefnitionsintroducedbyMller(2004)andBonabeau
Dessalles, Ferber, & Phan
andDessalles(1997),whichareoperativeinmulti-agentframeworksand
make sense from both cognitive and social point of view. A diagrammatic
4-Quadrantapproachallowsustounderstandcomplexphenomenaalong
both interior/exterior and individual/collective dimensions.
Introduction
The concept of emergence, frst discussed in philosophy, is also widely
used in complex adaptive systems literature especially in computer sciences
(Holland, 1998) and related felds (multi-agent systems, artifcial intelli-
gence, artifcial life...) as well as in physics, biology, and cognitive sciences.
Particular applications are the social and human sciences, and consequently
the design of artifcial society or agent-based computational economics
(ACE) framework by means of multi-agent systems (MAS). For instance in
a pioneering book on artifcial society and multi-agent simulations in social
sciences, Gilbert and Conte (1995) put the emphasis on emergence as a key
concept of such an approach: Emergence is one of the most interesting is-
sues to have been addressed by computer scientists over the past few years
and has also been a matter of concern in a number of other disciplines from biology
to political science (op.cit. p. 8). More recently, comprehensive discussion of
emergence issues can be found in Gilbert (2002) and Sawyer (2001a, 2004,
2005) for the social science and Sawyer (2002a) for the psychology. In eco-
nomics, ACE put also the emphasis on the question of emergence (see e.g.,
Axtell, Epstein, & Young, 2001; Epstein, 1999, 2006; Tesfatsion, 2002a,
2002b; Tesfatsion & Judd, 2006). In all these works, cognition and societies
are viewed as complex systems.
The present chapter discusses the impact of emergence on both downward
and upward effects, with applications to the social sciences. MAS allow us
to formalize in a single framework both bottom-up and top-down processes.
In multi-agent frameworks, properties of the whole system result from
the collective interactions between the parts (agents) by upward causation
(bottom-up process, compatible with methodological individualism); but,
to some extent, agents may be constrained by the whole top-down process,
compatible with holism or structuralism methodological point of view. This
downward effect may arise by means of the social dimension of beliefs
(Phan & Ferber, 2007) through the agents perception of social phenomena
or through structural properties of the agents social environment. Such a
downward determination is mainlybut not onlyassociated to cognitive
agents (Castelfranchi, 1998a-b). The process through which the macro-level
emerging social structure feedbacks into the micro-level by re-shaping
the elementary agents behaviors is also called immergence by (Gilbert,
1995, 2002).
This chapter provides in frst section a critical survey of emergence defni-
tions in literature and exhibits the common structure of the remaining issues.
Section 2 introduces and discusses the signifcance of formal defnitions of
emergence, with a special attention for those of (Mller, 2004; Bonabeau
et al., 1997). These formal defnitions are operative for modeling complex
artifcial societies using multi-agent oriented programming (Ferber, 1999)
and make sense from both cognitive and social point of views (Dessalles &
Phan, 2005). Complementary features related to complexity are introduced,
like detection and cognitive hierarchy. Finally, the last section proposes to
highlight the whole process of emergence in the cognitive and social con-
text using a comprehensive framework, the 4-Quadrant approach, which
allows an integrative understanding of complex phenomena at the light
of multi-agent oriented design in both an interior/exterior dimension and an
individual/collective dimension.
Some.Conceptual.Issues.on.Emergence
The notion of emergence has several meanings. In the vernacular language,
emergence denotes both a gradual beginning or coming forth, or a sudden
uprising or appearance; to emerge also means to become visible; for example,
emergence may denote the act of rising out of a fuid. This latter sense is
close to its Latin roots, where emergere is the opposite of mergere: to be
submerged. In what follows, we relate the act of rising out to the arising of
some phenomenon from a process, and note the fact that to become visible
presupposes some observer.
The common sense of emergence is therefore linked to the meaning of a pro-
cess that produces some phenomenon that might be detected by an observer.
In the feld of science, emergence was used by Newton in optics. By the 19
th

century, the word emergent was introduced into the felds of biology and
philosophy. In the latter, emergentism has a long history, from Mills chapter:
Of the composition of causes in System of Logic (Mill, 1843), Lewes dis-
tinction between resultant and emergent effects (Lewes, 1875); Morgan
(1923) and Broad (1925), to the contemporary debates about the philosophy
of mind around the mindbody problem. (For a synthesis, see among
others: McLaughlin (1992, 1997); Van de Vijver (1997); Emmeche, Koppe,
and Stjernfelt (1997); Clayton and Davies (2006); Kistler, 2006). Classical
defnition is given by Broad (1925):
The emergent theory asserts that there are certain wholes, composed (say)
of constituents A, B, and C in a relation R to each other; that all wholes
composed of constituents of the same kind as A, B, and C in relations of the
same kind as R have certain characteristic properties; that A, B, and C are
capable of occurring in other kinds of complex where the relation is not of
the same kind as R; and that the characteristic properties of the whole R(A,
B, C) cannot, even in theory, be deduced from the most complete knowledge
of the properties of A, B, and C in isolation or in other wholes which are not
oftheformR(A,B,C)(Broad,1935,Chapter2,underlinedbyus).
As underlined, British emergentism rejects reductionism: the properties of
the whole cannot be deduced from the properties of the parts. Several of
them consider emergence also from an ontological standpoint, coupled with a
layered view of nature. For ontological emergentism, the world is constituted
of hierarchically layered structures, or levels of organization. Lewes (1875)
places emergence at the interface between such levels of organization. Each
new layer is a consequence of the appearance of novel qualities, with an in-
creasing complexity. Emergent laws are fundamental; they are irreducible
to laws characterizing properties at lower levels of complexity, even given
ideal information as to boundary conditions. Since emergent features have
not only same-level effects, but also effects in lower levels, some speak of the
views commitment to downward causation (Campbell, 1974; also quoted
by OConnor et al., 2006). Although our main concern is downward causa-
tion, the philosophical position adopted here is more pragmatic, and deals
with epistemic emergentism more than ontological one. As a consequence,
our hierarchy of levels, and more generally our ontological commitment is
relative to a given epistemological stance.
Epistemic,.Ontological,.and.Methodological.Background
In numerous contemporary views of emergence (Kistler, 2006; OConnor &
Hong, 2006,), this concept of an epistemological category is referring to the
limits of human knowledge of complex systems. According to OConnor et
al. (2006), emergent properties are systemic features of complex systems
that could not be predicted (..) from the standpoint of a pre-emergent stage,
despite a thorough knowledge of the features of, and laws governing, their
parts. In addition, macroscopic patterns resulting from an emergent phenom-
enon could not be reduced
1
. In the past decades, a wide variety of defnitions
of epistemological emergence have been proposed. Although many of these
defnitions deal with non-reducibility, some of them are compatibles by some
ways with reducibility. As a consequence, the answer to the question: what
is an emergent phenomenon? depends on the concept of emergence one
invokes. A broad defnition of emergent property as been proposed by Teller
(1992): a property is emergent if and only if it is not explicitly defnable in
terms of the non-relational properties of any of the objects proper parts
(p. 140-141, underlined by us). This allows us to have some distance with
more canonic conditions of emergentism such as novelty, unpredictability,
and naturalistic hierarchy of layers. Then, it is possible to use the simple
two-level framework of organization (micro/macro) as only a methodologi-
cal one, linked with some epistemological stance. In this framework, Bedau
(1997, 2002) distinguishes two hallmarks of how macro level emergent
phenomena are related to their micro level bases:
1. Emergent phenomena are dependent on (constituted by, and generated
from) underlying process.
2. Emergent phenomena are (somehow) autonomous from underlying
process (Bedau 1997, p375).
Our approach of emergence is mainly locally epistemical, methodological,
and organizational (Van de Vijver 1997) (i.e., related with a specifc context
of knowledge and specifc tools (multi-agent systems). The corresponding
ontology is then methodologically driven, and relative to a specifc formal-
ism: the framework of knowledge of multi-agent oriented programming.
Furthermore, wedefneemergenceasaphenomenonrelativetoanobserver:
0 Dessalles, Ferber, & Phan
our concept of emergence is related to a particular framework of observation.
This point of view excludes all forms of Platonicism or other strong forms
of scientifc realistic commitment
2
: scientifc knowledge is not viewed as
the mirror of the nature. This epistemic point of view avoids numerous
questions addressed by the so-called orthodox emergentism in the debate
about non reductionist physicalism.
According to Van de Vijver (1997), from the organizational point of view on
emergence, the hierarchy of levels does not necessarily correspond to some
real hierarchy in the real world, but characterizes a locally relevant (from
some academic feld point of view for instance) conceptual organization of
the world (here artifcial world) in terms of:
1. An abstract closed system (the object of the study, or target system).
2. A discussion about the empirical/technical/conceptual relevance of such
a point of view.
3. A discussion about the relevance of the properties, ontologies, etc.,
related to the corresponding system (entities, levels of organization,
relations).
4. Both discussions must be related to the goal of the scientifc process.
In this pragmatic view, each type of explanation has its own goal, relevance,
and limits (Clark, 1996) and must be related to a specifc scientifc project. In
such an approach, both conceptual and operational models in generaland
multi-agent models in particularcan be viewed as mediators between theory
and empirical evidences (Morgan & Morrison, 1999). As such, models are
both means to and source of knowledge (Morgan et al., 1999, p. 35). Fur-
thermore, as Minsky (1965) said, to an observer B, an object A* is a model
of an object A to extend that B can use A* to answer questions that interest
him about A. Then, the model can be viewed as a specifc technology in the
process of learning and inquiry for knowledge, to answer questions. This
pragmatic point of view on the emergence discussed hereafter is consequently
contextual to the project of investigating complex social phenomena with
cognitive agents by means of both complex adaptive systems methodology
and multi-agent system modeling.
The.Varieties.of.Emergence:.Purpose,.Meaning,.and.Stakes
According to the usual characterization of complex systems, the properties
of the whole complex (social) system cannot be reduced to the properties
of the parts. It results therefore also from the relations between parts and, in
some cases, from some irreducible macro causal power from the whole (i.e.,
downward causation). We notice that the relational properties, which structured
the system, are neither at the level of the whole nor at the level of the parts,
while being constitutive of both. The nature (and possible reducibility) of
such a downward causation is one of the main debates in the feld of emergent
phenomena. As said previously, the answer to these questions depends on the
defnition of emergence one uses. Some authors have proposed to distinguish
different kinds of emergence, as for example nominal, weak, and strong
emergence for Bedau (1997, 2002), or weak, ontological, and strong
emergence
3
for Gillett (2002a-b). For Bedau, the broader (weaker) form of
emergence is called nominal. Nominal emergence concerns the existence
of some macro-property that cannot be a micro property. Each level has its
specifc distinct role and properties: macro-level emergent phenomena are
dependent on micro-level phenomena in the straightforward sense that wholes
are dependent on their constituents, and emergent phenomena are autonomous
from underlying phenomena in the straightforward sense that emergent prop-
erties do not apply to the underlying entities (Bedau, 2002). Under this latter
condition, strong emergence is the opposite of nominal emergence, as in this
case, emergent properties have irreducible causal power on the underlying
entities: macro causal powers have effects on both macro and micro levels,
and macro to micro effects are termed downward causation (Bedau, 2002)
Weak emergence is a subset of nominal emergence for which the emergent
phenomenon is not easy to explain, according to Simon: given the proper-
ties of the parts and the law of their interactions, it is not a trivial matter to
infer the properties of the whole (Simon, 1996, p. 184, quoted by Bedau,
2002) Accordingly, for Bedau (2002), weakly emergent phenomena are those
which need to be simulated, to be revealed: Assume that (a macro-state) P is
a nominally emergent property possessed by some locally reducible system
S. then P is weakly emergent if P is derivable from all of Ss micro facts but
only by simulation. According to the non-trivial dimension (surprising)
of emergent phenomena, the need for simulation seems to be a transitory
epistemic criterion only. If in a context of discovery, computer simulation
reveals some new emerging patterns, this is not a suffcient condition to have
no other way forever. Later justifcation by some explanatory formalism is
a possible outcome. Thus, a surprising (weak) emergent phenomenon could
become a simple nominal emergent one.
Stephan (2002a, 2000b) distinguishes different degrees in novelty, reducibility,
and predictability. He proposes an interesting discussion on the difference
between weak and strong forms of emergence in a larger framework, using
the difference between synchronic and diachronic emergentism (see
also Rueger, 2000a). The former postulates that a macroscopic emergent
phenomenon can be explained by the current (synchronic) interactions of the
interrelated microscopic entities. In other words, the center of interest is the
relationship between the interacting entities and the whole system resulting
from these entities and their relationship. In diachronic emergentism, on the
other hand, the emergent phenomenon occurs across time by means of se-
quential adaptation of microscopic entities. The center of interest is now the
evolution of both micro and macro structures, and not only the occurrence
of a particular structure. As underlined by Stephan (2002b) synchronically
emergent properties include also diachronically emergent ones, but not
conversely.
For instance, in a Wolfram one-dimensional network of automata (1984), a
specifc confguration of the network emerges at each step from the value
Figure 1. Emergence of Sierpinskys triangular structures within
the diachronic diagram of the Wolfram network of automata
SourceAmblard&Phan(2006,p.277;2007)
|
|
time
|.
structure of the network of automata..

of the automaton and the structure of their relations at the previous steps
(synchronic emergence). In some cases, identifed by both Wolfram and
Langton (1984), the existence of an attractor drives the system towards a
particular stable confguration (fxed point, cycle). In some others cases,
called by Langton (1989) the edge of Chaos, the evolution of the states of
the systems, from step to step, generates a particular structure, such as the
Sierpinskys triangular structures (Figure 1). This structure is only observable
from a diachronic perspective, and results from the succession of synchronic
emergence of macrostructures due to the local interaction of microstructures
(namely, the automaton) within the specifc one-dimensional nearest-neighbor
interaction.
For Stephan (2002a) the weaker version of emergentism (weak emergence)
can be characterized by three features. First, following the physical monism
thesis, only physical entities can bear the emergent properties or structures.
Secondly, emergent properties or structures are attributes of the system it-
self, and cannot be attributed to some systems part. Thirdly, the principle of
synchronic determination implies that all properties of the system nomologi-
cally depend on its micro-structures, namely, the parts and their relations.
Stephan underlines that this latter thesis of synchronic determination can be
understood as a stronger version of mereological supervenience. In mereo-
logical supervenience, the systems properties supervene on its parts and their
relations, but this does not imply their dependence on its micro-structures
(Stephan, 2002a, p. 80).
Numerous emergentists refer to debates about reductionism as well as about
the so-called mind-body problem, discussing in particular the notion of
supervenience introduced by Davidson (1970, 1980) and discussed by Kim
(1992, 1993, 1995, 1999) from the point of view of emergence. Superve-
nience is a relation that can be summarized in a slogan form, by the sentence:
there cannot be an A-difference without a B-difference. For application to
individually-related properties, this idea means that two individuals cannot
differ in M-properties without also differing in N-properties. The strongest
form (useful here) of supervenience asserts: A family of properties M
strongly supervenes on a family N of properties if and only if, necessarily,
for each x and each property F in M, if F(x) then there is a property G in N
such that G(x) and necessarily if any y has G it has F (Kim 1993, p. 65; see
also the distinction between weak and strong form of individual superve-
nience in the possible world modal framework, same book). According to
McLaughlin (1997), it is possible to defne emergence from this strong form
of supervenience: If P is a property of W, then P is emergent if and only if
(1) P supervenes with nomological necessity, but not with logical necessity,
on properties the parts of W have taken separately or in other combinations;
and (2) some of the supervenience principles linking properties of the parts
of W with Ws having P are fundamental laws (p. 39). The two important
features are the nomological (but non logical) necessity, and the notion of
fundamental law, which means that it is not metaphysically necessitated
by any other laws of W. As underlined by McLaughlin and Bennett (2006),
this defnition of emergence involves synchronic supervenience. This could
be problematic and requires at least a convenient concept of reduction (Kim,
1999, 2006) since if everyone agrees that reduction requires supervenience
(McLaughlin & Bennett 2006), the converse is false. In particular, non re-
ductive materialism rejects the reducibility argument and asserts that mental
properties are not reducible to physical ones. This kind of emergentist argu-
ments used in the mind-body debate has been re-used for the methodological
debate in the social sciences between holism and individualism.
Stephan (2002a, 2000b) claims that numerous formal approaches to com-
plex systems, connectionism, and cognitive science can be related to weak
emergentism. In this chapter, the possibility for MAS to encompass such a
limitation is precisely a major question of interest. This means that the strong
emergence directly addresses the questions of downward causation and reduc-
ibility. We do not discuss in this paper all the questions raised by Stephan.
When addressing the classical debate about methodological individualism
versus holism in the social sciences
4
, we do not need to discuss the question
of reduction to the material basis of human (and social) behavior.
Figure2.Irreducibilityandnoveltyinemergentphenomenon(adaptedfrom
Stephan2002b)
Synchronic
determination
Diachronic
determination
Weak
(reducible)
weak
emergentism
weak diachronic
emergentism
|
Irreductibility
Strong
(irreducible)
(strong)
synchronic
emergentism
strong diachronic
emergentism
..Novelety
Hence, we restrict ourselves to a very simplifed two-level framework where
individuals are the basic entities at the frst level, and where one wonders
how some social entities may have any existence. According to our dis-
cussion on the two dimensions of the so-called strong emergence (namely:
downward causation, reducibility), the nature of social facts will depend on
the effective class of emergence that we consider. From this limited point of
view, in order to summarize this two-level problem we introduce in Figure 2a
cross perspective on synchronic/diachronic-weak/strong emergence, adapted
from Stephan (2002b) without the case of unpredictability).
To sum up, while both downward causation and irreducibility are generally
considered by philosophers as necessary conditions for strong emergence,
the defnition of weak emergence depends on the author and remains unclear.
While irreducibility seems to be a necessary condition for British emergentism,
for Stephan (2002ab) reducibility corresponds to the case of weak emergence.
Then, relevant questions are: what is the criterion and framework of reference
to have a clear account of irreducibility and downward causation.
Irreducibility.and.Downward.Causation:.A.Synthetic.View
As weak emergence deals with upward causation and reductionism, (Bedau
2002; Gillet 2002b) relate emergence to the question of downward causa-
tion or macro-determinism. Strong forms of downward causation are
widely advocated by Sperry (1969, 1986, 1991) to deal with the mind-brain
interactions, and by Campbell (1974) to deal with hierarchically organized
biological systems. According to the downward causation, the behavior of
the parts (down) may be determined by some properties or behavior of the
whole (top). For instance, parts of the system may be restrained by some act
in conformity with the rules given at the system level. Causation would come
downward in conformity with a holistic principle rather than upward,
according to reductionism. The existence of irreducible downward causation
is sometimes used to discriminate between weak and strong emergence.
Bedau (2002) introduces a notion of weak downward causation in order
to keep the idea of emergent causal power from the whole upon the parts:
emergent phenomena without causal powers would be mere epiphenomena
(Bedau, 2002). Although we do not agree with the metaphysical commitment
of Bedau, his arguments in favor of weak downward causation are appre-
ciably persuasive. As pointed out by Kim (1992, 1999), the general form of
downward causation effects can be respectively incoherent, inconsistent with
the law of micro entities, and excluded by some micro effects. The exclu-
sion argument concerns the cases where the emergent macro causation can
be derived from the causal power of the micro constituents. This is effectively
the case for numerous emergent phenomena in complex adaptive systems
such as sand piles, snow avalanches and so on. In all models of statistical
mechanisms, the law of the emergent macro level can be derived from the
law of the elements and their interactions. As pointed out by Kim, weak
downward causation has paradoxical consequences. If these considerations
are correct, higher-level properties can serve as causes in downward causal
relations only if they are reducible to lower-level properties; the paradox is
that if they are so reducible, they are not really higher level any longer
(Kim 1999, p. 33). This problematic situation can be avoided by two means.
First, in all that cases, Bedau (2002) claims that there is no worry with weak
downward causation, because the sequence of causal effects is diachronic:
Higher-level properties can causally infuence the conditions by which they
are sustained but this process unfolds over time. The higher-level properties
arise out of lower level conditions, and without this lower level conditions,
the higher level properties would not be present. Second, both Bedau and
Kim acknowledge the relevance and the autonomy of the higher level from
a causal point of view in the case of an explanatory standpoint (Bedau) or a
conceptual interpretation (Kim): we interpret the hierarchy of levels of
concepts and descriptions, or levels within our representational apparatus,
rather than levels of properties and phenomena in the world. We can then
speak of downward causation when a cause is described in terms of higher-
level concepts, or in a higher level language, higher in relation to the concepts
in which its effects are represented. On this approach, then, the same cause
may be representable in lower-level concepts and languages as well, and a
single causal relation would be describable in different languages (Kim
1999, p. 33). This point of view is nothing but our organizational pragmatic
approach of emergence, a step in the process of knowledge.
Emergence.as.Perceptive,.Cognitive,.and.Social.
Phenomenon:.Looking.at.Social.Facts.in.a.New.Light
The respective role of social structures and individual action is a fundamental
issue in social theory. The relevant questions are does structure determine
action or action determine structure? Or is it a bit of both? (Hollis, 1994, p.
6). From the structural standpoint, individual action is externally constrained
by some holistic principles. For instance, the Rules of Sociological Method
of Durkeim (1895) is generally presented as a paradigmatic holistic point
of view, since social facts, taken as things, are external to individuals
and external observer have no direct access to those external things. On the
contrary, from the individualist standpoint, society is nothing but the result
of the individuals actions. Accordingly, for methodological individualism,
social phenomena must be stated by means of actions and interaction between
individuals. This does not mean necessarily that individuals are the only
relevant level for the social facts. Thus, Udehn (2001) distinguishes strong
methodological individualism, for which all social phenomena must be reduced
to the individuals behavior (e.g., Popper, 1966) and weak methodological
individualism, for which autonomous institutions and social structures can
shape the individuals behavior, even if social facts must be taken into ac-
count in individuals concepts.
Despite the commitment for the promotion of methodological individualism
by several sociologists (Boudon, 1998, 2006; Coleman, 1990; Elster, 1989,
1998), sociologists are often viewed as unrepentant holists. Then, accord-
ing to Granovetter (1985), the mainstream sociologists approach would be
over-socialized (related mainly to downward effects) while the economists
approach would be under-socialized (related mainly to upward effects).
Nevertheless, both approaches have been more sophisticated and are often
mixed (Hollis, 1994). Numerous scholars have proposed various forms of
sophisticated methodological individualism, and the non-reductive argu-
ment (of the social level to the individual) is a key condition for emergence.
Among these scholars, Sawyer (2001a, 2004, 2005) uses the notion of super-
venience to identify emergent social properties, which cannot be reduced
to an explanation in terms of individuals and their relationships. Sawyer
(2002b, 2003) calls non reductive individualism (NRI) this dualism at mid-
way between individualism and holism: NRI holds to a form of property
dualism in which social properties may be irreducible to individual proper-
ties, even though social entities consist of nothing more than mechanisms
composed of individuals (Sawyer, p. 266) According to the discussion on
non-reductive physicalism, Sawyer argues that both multiple realizability
and wild disjunction are necessary and suffcient conditions for emergence of
non reducible social properties. Multiple realizability appears when a single
social property can be generated by several micro-level mechanisms. This
argument is a common objection to methodological individualism (Kincaid,
1996, Zahale, 2003). Wild disjunction appears when these mechanisms are
not meaningfully related. Sawyers claims have been criticized. For Bunge
(2004), the notion of supervenience is less clear than the notion of emergence,
in particular in the case of diachronic emergence relative to a given system.
If emergence is defned as the rising out of a qualitative novelty, this new
property appears at some point in the development or the evolution of the
system. In contrast, supervenience does not use the concept of system and
levels of organization (Bunge, 2004, p. 377-78). This diachronic dimension
of social emergent has been underlined by Archer (1995) for whom social
structures emerged in the past from actions of agents; continue to exert effects
in the present. As a consequence the only pertinent concept of emergence is
diachronic (see also Manzo, 2007). As underlined previously, if the non-re-
ducibility argument is problematic in a synchronic context (and by extension
the related downward causation) this is not the case in a diachronic context.
Then, on the one hand, in a diachronic perspective with reifcation, non reduc-
tive emergence is less problematical. On the other hand, Sawyers criticism
against Archers arguments is less relevant in a socio-cognitive framework,
as suggested for instance by Castelfranchi (1998a, 1998b, 2000).
Bunge (1977b, 1979) proposes an individually-based systemic concept for
social analysis, in which both individual and collective take place. According
to Bunge, a systemic society is a system of interrelated individuals, i.e. a
system, and while some of its properties are aggregation of properties of its
components, others derive from the relationship among the latter (Bunge,
1979, p. 13-14). According to Lewes (1879) the frst relation (aggregation)
describes resultant effects, and characterizes methodological individual-
ism, and allows reductionism. The second relation is typically systemic,
and characterizes emergentism, where bottom up properties emerge from
the relations between the systems components, and are not possessed by
any component of the whole. More specifcally, Bunge defnes resultant and
emergent properties
5
as follows:
Let P p(x) be a property of an entity x S with composition C(x) {x}.
Then P is a resultant of hereditary property of x iff P is a property of some
component y C(x) of x other than x; otherwise, P is an emergent or Gestalt
property of x. That is:
1. P is a resultant of hereditary property of x =
df
P p(x) & (y)(y C(x)
& y x & P p(y))
2. PisanemergentorGestaltpropertyofx=
df
P p(x) & (y) (y C(x)
& y x & P p(y))(Bunge,1977b,Defnition2.16p.97).
Where is the basic symbol for negation. In addition, postulate 2.19 (Bunge,
1977b, p. 98) distinguishes reducible properties (eliminable in favor of micro-
based properties) from no-reducible but analyzable or explainable. Reducibil-
ity entails analyzability, but the converse is false. There is epistemological
novelty in the formation of attributes representing emergent (ontological)
novelty (id.). But the explanation of emergent phenomena does not involve
the elimination of ontological novelty. Bunges systemism can be analyzed
in both ontological and methodological perspectives.
In the ontological perspective, systemic collectivity is neither a set of indi-
viduals nor a supra individual entity transcending its members, but a system
of interconnected individuals. There are global properties, some of these
properties are resultant (reducible) some others are emergent from individual
interactions (non-reducible). For Bunge, systemic society cannot act on its
members, but members of a group can act severally on an individual. Finally,
social change is a change in the social structures of societyhence a change
at both societal and individual levels (Bunge, 1979, p. 16). The systemic
framework introduced by Bunge is an interesting frst step to encompass both
monist individualism and holism. But there are also some intrinsic limitations.
Signifcantly, several examples of Bunge are taken from natural science, not
from human and social sciences. Pure Bungian agents have limited cognitive
capacity and a lack of social intelligence (Conte, 1999). However, these
limitations also appear with many models of emergence in artifcial society
as underlined by Sawyer (2004, p 265). The emergence is viewed only as a
bottom-up process, without effective downward causation (see for instance
the paradigmatic models of Schelling (1969, 1978) and Axtell et al. (2001)
and their account by Dessalles et al. (2005) and Dessalles, Gallam, and Phan
(2006)). In the following, we introduce new formal frameworks that allow
us to encompass these limitations either with cognitive epistemic agents or
with less cognitive, behavioral ones. The following example suggests how
simple modifcation of information in the environment of individuals could
feedback from top to down, through a specifc mediator.
The case of a traffc jam, quoted by Bedau (2002) is a very interesting one
for the discussion of the macro to micro relationship. Traffc jams arise in
particular confguration of the micro constituents, and they are caused by
the composition of such micro determinants. The process of traffc jam is
then reducible in principle to (and can be simulated from) the behavior of
the basic entities (the cars) in a specifed environment. If we pay more at-
tention to the cognitive and social dimension of this problem, traffc jams
can arise because each individual does not have enough information in order
to be spontaneously (from the bottom up) coordinated with the others in
the use of limited capacity of traffc. In some highway infrastructures this
worry has been encompassed by the use of an external information system
that transmits messages to the car drivers about traffc. This results often in
a better coordination and in the decrease in traffc jams, as the drivers do
not act myopically, but take into account this information on the possible
occurrence (emergence) of a traffc jam at the macro level to modify their
own behavior at the micro level.
This example underlines the fundamental difference between complex in-
teractive systems with reactive agents and systems with cognitive ones. In
the former, in the case corresponding to weak emergence all the causality is
reducible to the micro elements. In the latter, the existence of some social
mediator, able to support feedback effect from the macro level to the micro
can develop some autonomous properties and causal effects upon the micro
behavior that cannot be directly reducible to micro causation. This could be
modeled by means of reifcation of these social mediators, which makes sense
from a social point of view (Phan et al., 2007, and section 3 on 4-Quadrants
below). The generation of such mediators, from the bottom up can be viewed
as a weak emergent phenomena, but the reifcation of some social repre-
sentation of this emergent pattern and the feedback effect from this social
object to the individual behavior can be viewed as a qualitative change in the
model (or a shift from a model to another, at a different level, according to
Bonabeau et al. (1997), that is certainly the strongest form of emergence than
those addressed by Bedau. Indeed, in the larger new qualitatively different
model, the unsurprised re-emergence of a previously reifed phenomenon
can be viewed as a weak emergent phenomenon on the one hand; but from
the point of view of the lower level modelbefore reifcationthat is a
strong emergent one. The following formal defnition of emergence allows
us to explain this question more precisely. In this chapter, we do not address
these questions directly, as we limit ourselves to discussing social behaviors
in artifcial societies; but the opposition downward versus upward causation
proves to be a central one in the feld of social sciences.
Formal Defnitions of Emergence
The present chapter is an attempt to integrate them into one single framework,
in which the whole is a collective of cognitive agents (according to meth-
odological individualism), while the agents are to some extent constrained
by the whole (downward causation), by means of the social dimension
of their belief, their perception of social phenomena, or by some structural
properties of the collective as well
6
. For this purpose, we rely on the dis-
tinction introduced by Labbani, Mller, and Bourjault (1996) and Ferber,
Labbani, Mller, & Bourjault (1997) and developed by Mller (2004) in the
feld of multi-agent systems, between weak and strong emergence. The
latter refers to a situation in which agents are able to witness the collective
emergent phenomena in which they are involved, which opens the road for
both upward and downward causation.
In ACE (Epstein, 2006; Tesfatsion & J udd, 2006) and computational social
sciences (Gilbert, 2007), emergence is strongly related to the Santa Fe Ap-
proach to Complexity (SFAC). In accordance with descriptive emergentism,
SFAC calls emergence the arising at the macro level of some patterns, struc-
tures and properties of a complex adaptive system that are not contained in
the properties of its parts. Interactions between parts of a complex adaptive
system are the source of both complex dynamics and emergence. An interest-
ing part of the emergence process concerns the forming of some collective
order (coherent structures or patterns at the macro level) as a result of the
agents interactions within the systems dynamics. For the observer (i.e. the
computational social scientist) this collective order makes sense by itself and
opens up a radically new global interpretation, because this does not initially
make sense as an attribute of the basic entities.
In this chapter, our concern is about formal models of emergence in MAS
with cognitive and social agents. Therefore, we deal mainly with formal
defnitions of emergence, operative for MAS. Formally, in MAS, emergence
is a central property of dynamic systems based upon interacting autonomous
entities (agents). As previously mentioned, the knowledge of entities attri-
butes and rules is not suffcient to predict the behavior of the whole system.
Such a phenomenon results from the confrontation of the entities within a
specifc structure of interaction, which is neither at the level of the whole
system nor at the level of the entities, but constitutive of both. Accordingly,
a better knowledge of the generic properties of the interaction structures
would make it easier to have better knowledge of the emergence process (i.e.,
morphogenetic dynamics). From this point of view, to denote a phenomenon
as emergent does not mean that it is impossible to explain or to model the
related phenomenon. For this reason Epstein, 1999) uses the word genera-
tive instead of emergent in order to avoid a philosophical debate about
emergence.
Some Defnitions of Emergences.in.Complex.Systems
Various attempts have been made to defne emergence in an objective way.
Some defnitions refer to self-organization (Varela, Thompson & Rosch,
1991), to entropy changes (Kauffman, 1990), to non-linearity (Langton,
1990), to deviations from predicted behavior (Cariani, 1991a; Rosen, 1977,
1978, 1985) or from symmetry (Palmer, 1989). Other defnitions are closely
related to the concept of complexity (Bonabeau, Dessalles, & Grumbach,
1995a, 1995b; Cariani, 1991b; Kampis, 1991a, 1991b).
In statistical mechanics (Galam, 2004), as well as for models in economics or
social sciences having the same structure than models of statistical mechan-
ics
7
, emergence may be related to an order parameter which discriminates
between at least two phases, each one with a different symmetry associated
respectively to a zero and a non-zero value of the order parameter. Each
problem has its own specifc order parameter
8
.
For instance, in the Ising model, where individual spins take their values in
{- 1, +1}, the order parameter is the magnetization M, given by the sum of
all spin values divided by their total number. When M=0, the state is para-
magnetic (i.e., disordered in the spin orientations) while long range order
appears as soon as M 0. A majority of spins are then oriented to either -1
or +1, and an order is likely to emerge. Two ordered phases are thus possible
in theory, but only one is effectively achieved. The order parameter provides
a signature for the emergent phenomenon.
Although these defnitions make use of concepts borrowed from physics
and information science, they all involve inherently contingent aspects, as
the presence of an external observer seems unavoidable. Even a change in
entropy supposes that an observer be able to assess the probability of vari-
ous states.
Emergence.as.a.Phenomenon.Related.to.an.Observer
The unavoidable presence of an observer does not preclude, however, the
possibility of extending the defnition of emergence to include non-human
observers or observers that are involved in the emerging phenomenon. In our
quest for strong emergence, we wish to assign the role of the observer to
elements of the system itself, as when individuals become aware of phenom-
ena affecting the whole society. This kind of self-observation is only possible
because what is observed is a simplifed state of the system. Emergence deals
precisely with simplifcation.
Ronald and Sipper (2001) introduce a new approach called emergent engi-
neering in order to have a controlled concept of the previously mentioned
concept of surprise. This approach opposes the classical engineered auto-
mation, based on unsurprising design, and the biologically inspired automa-
tion system, which allows the possibility of unsurprising surprise. Many
engineered emergent systems are based on this concept (e.g., Vaario, Hori,
& Ohsuga, 1995). We do not deal directly with emergent engineering, but we
discuss the framework used by this author based on a specifc formal test of
emergence, previously presented in Ronald et al. (1999). This test of emergence
involves two functions, which can be assumed by the same individual or by
two different persons: (1) a system designer and (2) a system observer. An
emergent phenomenon can be diagnosed by combining the three following
conditions (Ronald et al., 2001, p. 20)
1.. Design: The system has been constructed by describing local elementary
interactions between components (e.g., artifcial creatures and elements
of the environment) in a language L
1
.
2.. Observation: The observer is fully aware of the design, but describes
global behavior and properties of the running system, over a period of
time, using a language L
2
.
3.. Surprise: The language of design L
1
and the language of observation
L
2
are distinct, and the causal link between the elementary interactions
programmed in L
1
and the behaviors observed in L
2
is non-obvious to
the observer- who therefore experiences surprise. In other words, there
is a cognitive dissonance between the observers mental image of the
systems design stated in L
1
and his contemporaneous observation of
the systems behavior stated in L
1
.
The question is then how easy it is for the observer to bridge the gap between
L1 and L2. The authors use artifcial neural network classifers to evaluate this
gap. Within this framework, an unsurprising surprise can be defned as an
expected surprise. This question is exemplifed later, within the (Bonabeau
et al., 1997) framework of emergence as reduction of complexity within the
observation system.
The framework of Ronald et al. (1999) and Ronald et al. (2000) together with
Forrests defnition of emergent computation (Forrest, 1990) allow Mller
(2004) to defne emergence in SMA as occurring between two organization
levels, distinguishing the process itself and the observation of that process.
The process concerns the evolution of a system formed by entities in interac-
tion using a language L
1
. These interactions may generate observable epiphe-
nomena. At the observation level, epiphenomena are interpreted as emerging
through specifc calculation using another language L
2
. Finally, emergence
is defned as a particular relationship between the two languages where L
2
is not compositionally reducible to L
1
in the sense of Bunge (1977a). For
Mller, weak emergence arises when the observer is external to the system.
This account is stronger than the notion of weak emergent phenomenon in
the sense of Bedau (1997, 2002) by adding to the necessity of simulating, the
intrinsic irreducibility of the two description languages. Strong emergence
arises when the agents involved in the emerging phenomenon are able to
perceive it. In this latter confguration, the identifcation of epiphenomena
by agents interacting within the system involves a feedback from the obser-
vation to the process. There is a coupling between the process level and the
observation level through the agents because the agents are using both L
1
and L
2
. This form of strong emergence is thus immanent in such a system. In
order to avoid misinterpretation, we call M-Strong the strong emergence
in the sense of Mller (2004).
To summarize, if there is M-Strong emergence, the system becomes refexive,
through the mediation of the agents.
1. Agents are equipped with the capacity to observe and to identify a
phenomenon in the process, which represents the evolution of the sys-
tem in which they interact. This capacity of observation and the target
of such observation must then be suffciently broad to encompass the
phenomenon as a global one.
2. The agents describe this epiphenomenon in a language other than the
one used to describe the process
3. The identifcation of an emergent epiphenomenon by the agents in-
volves a change of behavior, and therefore a feedback from the level of
observation to the process.
This category of m-strong emergence is important to model artifcial societies
(Castelfranchi, 1998a, 1998b; Gilbert, 1995). This is the case also even if
there is a mix of strong and weak emergence in most multi-agent based social
simulation (Drogoul & Ferber, 1991; Drogoul, Ferber, & Cambier, 1994).
Learning.and.Intrinsic.Emergence
Crutchfeld (1994), Bersini (2004), and Philemotte & Bersini, 2005a, 2005b)
propose to consider an alternative defnition of emergence, called intrinsic
emergence. They suggest to characterize emergence as an autonomous
increase in the systems computational capabilities. Such a defnition is
supposed to be more objective, as a natural way to avoid the presence of
an external observer in charge of detecting emergence. (Philemotte et al.,
2005a) implemented a situation of intrinsic emergence. In their system, a
cellular automaton is evolved through a genetic algorithm (GA) until it is
able to perform some arithmetic operations on a limited set of operands. As
usual for cellular automata, the rules, which for each cell, decide of its next
state, take as input the previous state of neighboring cells. In Philemotte et
al.s system, a second genetic algorithm is in charge of simplifying inputs for
each cellular automaton by limiting the number of neighboring cells actually
taken into account, so as to make the learning task easier for the frst GA.
Intrinsic emergence is claimed to occur whenever the second GA is able to
isolate a relevant portion of the neighboring input and thus to signifcantly
improve the learning effciency of the overall system. Philemotte and Ber-
sini were able to observe such sudden improvements when the two genetic
algorithms cooperate.
Emergence.as.a.Complexity.Drop
In Bonabeau et al. (1997), emergence is defned as an unexpected complex-
ity drop in the description of the system by a certain type of observer. Such
a defnition is claimed to subsume previous defnitions of emergence, both
structural (dealing with levels of organization) and epistemological (dealing
with deviation from some models predictions). In each case, the observer is
able to detect a structure, such as the presence of relations between parts of
the system, or some form of behavior like a characteristic trajectory. Struc-
tural emergence occurs whenever a collection of similar elements turns out
to be more structured than anticipated. This augmentation of structure can
be characterized by a decrease of complexity.
E = C
exp
C
obs
(1)
Here, E stands for the amplitude of the emergence; C
exp
is the expected
structural complexity and C
obs
the structural complexity actually observed.
Structural complexity is defned as the algorithmic complexity relative to a
given set of structural descriptors. In order to use algorithmic complexity
to describe fnite systems, we abandon the generality of the concept as it
was defned by Kolmogorov, Chaitin, and Solomonov (Li & Vitanyi, 1993),
considering that the description tools are imposed by the observer and not
by a generic Turing machine. We defne the relative algorithmic complexity
(RAC) of a system as the complexity of the shortest description that a given
observer can give of the system, relative to the description tools available
to that observer. Emergence occurs when RAC abruptly drops down by a
signifcant amount.
For our purpose here, we must restrict the defnition. We consider a specifc
class of observers in order to get closer to what human observers would
consider as emergence. Following Leyton (2001), we impose the observers
description tools to be structured as mathematical groups. The observer may
be considered as being a Leyton machine, for which any structure is ob-
tained through a group-transfer of other structures (Leyton, 2001). Any level
of organization that can be observed has operational closure and is structured
as a group, and the only structures that can be observed are the invariant of a
group of operations. Moreover, the observer is supposed to have hierarchical
detection capabilities. This means that all elements of the system that the
observer can consider have themselves a group structure.
For structural emergence to occur, it is important that there be an unexpected
complexity decrease. This may happen either because the detection of the
higher structure was delayed, as when one needs time to recognize a Dal-
matian dog in a pattern of black and white spots. It may also happen when
adding a new observable that, instead of increasing the overall complexity
of the system for the observer; it paradoxically decreases it (Bonabeau et
al., 1997).
Emergent phenomena are naturally described in two-level architecture (Figure
3). In such a framework, objects at the two levels make sense only because
some observer is able to detect them. The detected object at the upper level is
composed of objects of the frst level. Correspondingly, the upper level detector
is triggered by the activity of lower level detectors. The systems complexity,
defned as the minimal description that can be given of its state, drops down
by a signifcant amount when an upper-level detector becomes active, as its
activity subsumes the activity of several lower-level detectors.
Let us call s the emerging phenomenon, {d
i
} the set of lower-level detec-
tors and D the higher-level detector. Before emergence occurs, the expected
complexity may be written:
C(s & {d
i
}) =
i
C(d
i
) + C(s|{d
i
})
The notation C(a|b) means the complexity of a when the description of b is
available. If s designates a pattern of black and white patches, the{d
i
} may
refer to the detection of black patches. In this case, C(s|{d
i
}) is zero, as the
scene is entirely described once the{d
i
} are. Let us suppose that a new detec-
tor is taken into account. The expected complexity becomes:
C
exp
= C(s & D & {d
i
})
binary sensors objects
concepts
Q1
P1
P1
P1
P1
P1
P1

Figure3.Parallelismbetweenhierarchies:levelofdescription,levelofob-
servations (detectors), and conceptual level (association concepts-detectors)
(source:Dessallesetal.,2005)
Suppose the scene is described using D frst. Then, the actual complexity
becomes:
C
obs
= C(D) +
i
C(d
i
|D) + C(s|D & {d
i
}) (2)
Most of the time, C
obs
=C
exp
, which means that the complexity of the new
detector compensates what is gained by using it. In our example, dividing
the pattern into four regions and describing each region in turn would pro-
vide no complexity decrease. If, however, D subsumes some of the d
i
, then
C(d
i
|D) becomes small or even zero, and C
obs
gets smaller than C
exp
. This is
when emergence occurs.
In our example, D may be the shape of a Dalmatian dog. Many of the black
and white patches become predictable as soon as the dogs shape is recog-
nized. This sudden upper-level pattern recognition decreases the overall
complexity according to the preceding formula, giving rise to computable
amplitude of emergence.
Note that formulas (1) and (2) make a prediction that is not acknowledged
in most models of emergence. The emerging characteristic must be simple.
The simpler it is, the more signifcant is the emergence. In formula (2), it
is important that C(D) be small, as a large value would ruin the emergence
effect. In our example, a Dalmatian dog constitutes a familiar shape that has
therefore low complexity.
This requirement that the emerging property be simple seems to be verifed
in all examples to be found in the literature. This statement may be surprising
at frst sight. On certain occasions, emergence seems to involve an increase
rather than a decrease of complexity. Examples such as phase transition or
bifurcation into chaos come to mind. In such cases, however, what is remark-
able and simple is not the resulting state, but the point of bifurcation. If phase
transitions were fuzzy (e.g., if the transition from water to ice was progres-
sive between +10C and 10C) emergence would be much less obvious.
We note also that taking a higher-level detector D into account undoubtedly
makes things more complicated, as it increases the observational hierarchy.
This price paid to complexity is taken into account by the term C(D) in (2).
Emergence only occurs when this term is more than compensated by the low
value of the other terms of Cobs.
Relationship.with.Others.Concept.of.Emergence
We may wonder how the preceding defnition of emergence as a complexity
shift relates to other defnitions reviewed in this chapter. As shown in Deguet,
Demazeau, and Magnin (2006), the change of description language invoked
by Mller or by Ronald et al. amounts to taking new detectors into account.
This language change is captured by D in the preceding formula. The non-
obvious character of the behavior described in the upper-level language, as
invoked by Ronald et al., corresponds in our framework to the unexpected
complexity shift.
Philemotte and Bersinis notion of intrinsic emergence also relates to the
previous defnition. Their defnition is original, and is not limited to the
description of structural patterns. We may call it behavioral emergence, as
the criterion for emergence is a discontinuity in performance rather than a
discontinuity in structural complexity. We may, however, ask what is emerg-
ing in Philemotte and Bersinis two-level GA-based cellular automaton.
If the general defnition of intrinsic emergence is restricted to describing
some discontinuity in effciency, then the answer is that nothing emerges.
In their particular experiment, however, a relevant input flter can be said to
emerge. For some defnition of complexity, indeed, intrinsic emergence is
well described by defnition (1). The measure of complexity to be considered
here is the size of the relevant search space. When systematically ignoring
a portion of the input, the second GA dramatically reduces the space where
the frst GA will fnd an effcient rule for the cellular automaton. This presup-
poses, however, that the input flter does not exclude convenient solutions. If
complexity is set to a maximal value when no adequate rule is learned, then
intrinsic emergence can be said to correspond to a complexity drop. Note,
however, that intrinsic emergence, contrary to structural emergence, does
not rely on the complexity of structure (e.g., the complexity of hierarchical
group structure, but relies on learning effciency which directly correlates
with the size of the fltered search space).
Defnition (1) may be also applied to cases of diachronic emergence. The
fact that a given structure can only be detected by comparison between suc-
cessive states of the system may be merely ignored when considering com-
plexity shifts. Structure and thus unexpected simplicity is discovered in the
set of successive time slices. Diachronic emergence, according to defnition
(1), occurs whenever the complexity of this set turns out to be simpler than
anticipated.
Quadrants:.An.Integrative.View.of.
Multi-Agent.Systems
In order to give a comprehensive view of emergence in MAS it is necessary
to understand the various perspectives and components that make a MAS,
and thus to use an integrative view of MAS (Ferber, 2007, 2007b, Phan et al.,
2007), which is inspired from those of Wilbert (2000). This diagrammatic
framework is designed in order to provide a two-dimensional heuristic de-
scription of the complex relationship within social systems.
The.4-Quadrant.Framework
The 4-Quadrant approach resides in a decomposition into two axis: individual
vs. collective perspectives on the one hand, and interior (i.e. mental states,
representations) vs. exterior (i.e., behavior, objects, organizations) perspec-
tives on the other hand. These two axis taken together provide a four-quadrant
map where each quadrant must be seen as a perspective in which individuals,
situations and social systems, as well as the architectural design of artifcial
society may be described and discussed as it is shown on Figure 4.
The upper half of the diagram is related to the individual aspects of the MAS
(i.e., agents), whereas the lower half is dedicated to its collective aspects
(i.e., societies of various form and size). The left half is related to the interior
aspects, which reside only in the view of agents, and the right half is about
exterior (i.e., manifestations of the behavior and traces in the environment,
which may be seen by an outside observer). The I-I (interior-individual, upper
left) quadrant, is about emotions, beliefs, desires, intentions, of an individual
(i.e., about its mental states, its subjectivity). The E-I (exterior-individual,
upper right) quadrant describes physical bodies, concrete objects, and also
behaviors of individuals. The I-C (interior-collective, lower left) is about
shared knowledge and beliefs, collective representations, ontologies, social
norms, and represents the inter-subjective part of individuals, what could
be called the noosphere. The E-C (exterior-collective, lower right) is about
material or formal social structures such as institutions and organizations (i.e.,
collective forms and structures of groups and systems) what could be called
the sociosphere. According to this decomposition, it is clear that emergence
may appear either on the internal side or on the external side. If emergence is
seen as a construction going from the individual to the collective level, and
downward causation as constraints going from the collective to the individual
level, the 4-Quadrant map shows that emergence may appear either on the
internal or the external side. On the internal side, collective representations,
general concepts and ideas, arise from external beliefs and goals. As such
elements of the noosphere, which result from the composition of individual
representations and beliefs, act as constraints for the beliefs, objectives, and
way of thinking of individual agents. It is as if things could only be thought
through the paradigms and representations of the collective level. On the
external side, social structures, which result from the activities of agents,
act as constraints for their possible behavior.
Representation of Weak Emergent Phenomena
Fundamental questions about the emergence properties in weak emergent
phenomena need to be explained. We claim that the presence of an external
observer being able to discern an emergent phenomenon and level of orga-
Figure4.The4-Quadrantmap(adaptedfrom:Ferber2007,2007b,Phan
etal.2007)
Internal-Individual (I-I)
I Subjectivity
<mental states, emotions, beliefs
desires, intentions, cognition>
Interiority
External-Individual (E-I)
It, This Objectivity
<agent behavior, object, process,
physical entities >
Observables, exteriority
Internal-Collective (I-C)
We Inter-Subjectivity
<shared / collective knowledge
invisible social codes and implicit
ontologies, informal norms and
conventions>
Noosphere
External-Collective (E-C)
Them, All This Inter-Objectivity
<reifed social facts and structures,
Organizations, institutions>
SocioSphere
nization is unavoidable. Accordingly, who is this observer? From the point
of view of social sciences, what does the higher level of organization consist
in? For whom does this level make sense?
To understand the emergent phenomenon, we need to introduce an observer
(i.e., another agent), which sees a multi-agent system from an outside position.
In the weak form of emergence, the observer is an agent that stands outside
of the system, and thus outside of the four quadrants (Figure 5). As such, its
observation shows a reduction of complexity when the system is seen from
the E-I quadrant or from the E-C quadrant. The arrow in the fgure between
these two quadrants shows both a new structural pattern of organization,
which arise from interactions of the individual level, and the conceptual
simplifcation, which comes from a more abstract level of analysis.
An example of weak emergence is given by aggregation mechanisms, such
as the one described by Schelling in its model of segregation (Schelling
1969, 1978). Schellings aim was to explain how segregationist residential
structures could spontaneously occur, even when people are not segregation-
ist themselves. The absence of a global notion of segregationist structures
(like the notion of ghettos) in the agents attributes (preferences) is a crucial
feature of this model. Agents have only local preferences concerning their
preferred neighborhood, but the play of interactions generates global segre-
gation (Figure 6).
Agents choose their area in relation to the colors of their neighborhood.
Though agents may be weakly segregationist (each agent would stay in a
neighborhood with up to 62.5% of people with another color), segregation
Figure 5. Weak emergence (emergence of structure) seen from an (external)
observer point of view
Behavior
of agents
Structures
???
Observer

occurs spontaneously: in Figure 6a, no agent wishes to move, but this is an
unstable equilibrium. A slight perturbation is suffcient to induce an emer-
gence of local segregationist patterns (Figure 6b and 6c).
Local interactions are suffcient to generate spatial homogeneous patterns.
Spatial segregation is an emerging property of the systems dynamics, while
not being an attribute of the individual agents. Sometimes, local integrated
(non-homogeneous) patterns may survive in some niches. But such integrated
structures are easily perturbed by random changes, while homogeneous struc-
tures are more stable (frozen zones). Complementary theoretical developments
on Schellings model of segregation can be found in the growing literature
on this subject (see Dessalles et al., 2005 for further references). Indepen-
dently of the question of the empirical relevance of Schellings model, this
pioneering work is generally viewed as a paradigmatic example of the frst
generation of agent-based models, producing macro-social effects from the
bottom-up (Amblard & Phan, 2007). Figure 7 represents Schellings model
in the 4-Quadrant perspective: individual behaviors (E-I quadrant) based on
simple preference choices (I-I quadrant) result in a global pattern, the emer-
gence of ghettos (E-C quadrant) as an external observer (e.g., researcher,
experimentalist) could seen.
Figure 6. Original (checkerboard) Schelling Model (Source : Source:
http://www-eco.enst-bretagne.fr/~phan/complexe/schelling.html and Phan
(2004a))
(a) fully integrated
population equilib-
rium
(b) discontented
agents are crossed
(c)convergenceafter4
iterations
M-Strong.Emergence.from.the.4-Quadrant.Point.of.View
Among fundamental questions raised by emergence in the social feld, the
problem of the existence of some social entities (or social objects) is of
great importance for the modeling of artifcial society, in particular for the
so-called immergence (Gilbert, 1995).
M-strong emergence, from the standpoint of Mller, can be seen as a process
in which agents are able to observe and consider the situation from both an
individual and collective point of view. M-strong emergence arises when
global structures and/or abstract entities are considered by individuals to
determine their individual behavior.
In Schellings model, agents do act according to a simple preference rule,
which only considers the status of their neighborhood. Let us suppose now
that they possess a sense of membership. Each agent (grey or black) thinks
that it is a member of a community (grey or black) and that newcomer must
join their community and live close to agents of their own color. Let us
also suppose that there is some kind of penalty if an agent does not follow
that rule. Situations as this one is characteristic in human population, see
for instance the Capulet and Montaigu confict in Romeo and Juliet. This
(I-I)
(I-C)
(E-C)
(E-I)
A ghetto as a
simple set of agents
(I-I)
(I-C)
(E-C)
(E-I)
(I-I)
(I-C)
(E-C)
(E-I)
A ghetto as a
A ghetto as a
Figure7.Themethodologicalindividualistapproach:Weakemergenceof
ghettos in Schellings model of segregation
restriction imposes a downward causation (or immergence) relation: each
agent is constrained by its membership to its own community, the global level
reacting back to the agent level. A new agent has to choose its own camp
and cannot go freely to a specifc location.
Segregation and ghettos will appear, but this time as a result of explicit mem-
bership internalized by agents. The overall situation may be depicted as in
Figure 8, where each agent has a representation of being a member of its com-
munity. The grey and black community is then reifed in the I-C quadrant
as collective ideas and concepts, resulting in an m-strong emergence.
From the 4-Quadrant point of view emergence is not a straightforward
transition from the individual to the collective level: all quadrants are
deeply interconnected. Thus, an important issue in emergence would be to
dynamically relate weak and m-strong emergence using the interconnection
of the four quadrants. We have seen that Schellings model produces a weak
emergence, but that agent behaviors based on community membership result
in an m-strong emergence. But very often, in empirical social phenomena,
both emergences occur. First a weak emergence appears producing a collec-
tive structure, which is then observed by agents. These observations, added
to communications between agents, produces collective ideas or concepts,
which are then used by agents for their subsequent behavior, resulting in an
M-strong emergence.
Figure8.M-Strongemergence(emergenceofstructure)seenfrombothan
agent and an (external) observer point of view
Behavior
of agents
Collective
structures
???
Observer
Collective
ideas

If we trace the dynamics of the whole process, we have the following cycles
of dependence, which take place between the four quadrants:
a.. I-I..E-I: Transformation of individual representations into individual
behaviors in (E-I) (only given here as a bootstrap process)
b.. E-I..E-C: Weak emergence as creation of new collective patterns in
(E-C)
c.. E-I..E-C..I-I: Cognitive individual observe collective patterns in
(E-C) and subjectively represent or categorize individuals behavior
(E-I) by subsumption under some collective category or structure in
(I-I).
d.. I-I..E-I..E-C..I-C: Emergence of new collective (public) repre-
sentation (idea, concept, category...) through communication between
agents driving to reifcation of individuals beliefs in (I-C).
e.. I-I..I-C..E-C..E-I: Individual behaviors in (E-I) constrained by
individual (I-C) and collective representation (I-C) and by social and
organizational structures (E-C) as well.
f. Go to (b) in a recursive loop, in order for the whole process to con-
tinue.
Individuals mind is at the origin of their behavior. Interactions between
these behaviors produce emergent phenomena at the collective level (a).
These behaviors together will possibly result in a collective structure (weak
emergence) (b). Then, if agents have a suffcient level of cognition to be
able to observe and represent they will subjectively categorize individuals
behavior by subsumption under a category or structure within the internal-
individual quadrant (I-I) (c). Furthermore, if these agents can communicate
about these categories / representations, then a new collective representation
(idea, concept, category...) will appear in the interior-collective quadrant
(I-C) (d) (Steels 2006).The collective representation will then be used to
constrain the agent behaviors, in a downward causal link (e). The whole
process continues with (b) thus forming a loop from the external individual
level (E-I) to the external collective level (E-C) and back to E-I, through the
various individual and collective representations.
Conclusion
The refexivity meditated by the agents consciousness and/or aware-
ness appears to be a determinant characteristic that distinguishes systems
involving human agents from systems made of non-conscious or material
entities. Within agents, it is interesting to distinguish a hierarchy in the cog-
nitive capacity of agents, from reactive agent to epistemic agent
9
(Bourgine,
1993; Phan, 1995). If a reactive agent cannot be considered as an observer
of its environments, both behavioral and epistemic agents have the cognitive
capability to process available information and can be viewed as observers
of the process in which they take part. In this process a behavioral observer
only takes into account some visible characteristics of its environment, while
an epistemic observer models and simulates in some way this process. Ac-
cordingly, behavioral and epistemic agents can contribute to strong emergence
through consciousness. By contrast, a reactive agent has no consciousness
and contributes only indirectly to strong emergence, which is mediated by
the environment (Labbani et al., 1996). The general socio-cognitive process
briefy introduced here should be augmented and detailed to form a better
understanding of the m-strong emergence process which arises in complex
social systems. Emerging phenomena in a population of agents are expected
to be richer and more complex when agents have enough cognitive abilities
to perceive the emergent patterns or when the structures of the collective
can detect emergent phenomena and feedback on the agents level. Such
feedback loops between emerging collective patterns and their individual
components allow us to have more sophisticated design of agents in artifcial
societies. This requires complementary developments, like those discussed
in Phan et al. (2007) on the ontological status of social belief, or like the
notion of social intelligence presented by Conte (1999) as a property of
socially situated agents, and more generally like in all the works initiated by
Castelfranchi and Conte and co-authors since 1995.
Acknowledgment
The authors acknowledge J ean Pierre Mller for valuable remarks and others
intellectual contributions, Alexandra Frnod for signifcant corrections on the
original draft, as well as the ANR program Corpus for fnancial support of
the project CO-SMA-GEMS. Finally, we thank also anonymous referees
for valuable remarks on a previous version of this chapter. Denis Phan is a
CNRS member.
References
Amblard, F. & Phan, D. (2007). Multi-agent models and simulation for
social and human sciences. Oxford: The Bardwell-Press. Forthcoming
in September, translated from part of: Amblard F. Phan D. dir. (2006).
Modlisation et simulation multi-agents, applications pour les Sciences
de lHomme et de la Socit, Londres, Hermes-Sciences & Lavoisier,
with new Chapters).
Archer, M. S. (1995). Realist social theory: The morphogenetic approach.
NY: Cambridge University Press.
Archer, M. S. (1998). Social theory and the analysis of society, in May &
Williams (Eds.) Knowing the Social World, Buckingham-Open Uni-
versity Press.
Axtell, R., Epstein, J . M., & Young, H. P. (2001). The emergence of classes
in a multi-agent bargaining model. In Durlauf, & P. Young (Eds.), Social
dynamics, (pp. 191-212). Cambridge, MA: The MIT Press.
Baas, N. A. (1994). Emergence, hierarchies, and hyperstructures. In C. G.
Langton (Ed.), A life III, Santa Fe Studies in the Sciences of Complexity,
Proc. (Vol. XVII, pp. 515-537. Redwood City, CA, Addison-Wesley.
Baas, N. A., & Emmeche, C. (1997). On emergence and explanation. Intel-
lectica,25, 67-83
Beckermann, A., Flohr, A. H., & Kim, J ., (1992). Emergence or reduction?
Essays on the Prospects of Nonreductive Physicalism. Berlin & New
York: Walter de Gruyter.
Bedau, M. A. (2002). Downward causation and the autonomy of weak emer-
gence. Principia, 6-1 J une, special issue on Emergence and Downward
Causation, 5-50.
Bedau, M. A. (1997). Weak emergence, Nos, 31, Supplement: Philosophical
Perspectives, 11, Mind, Causation, and World, 375-399.
Bersini, H. (2004). Whatever emerges should be intrinsically useful, in Ar-
tifcial life 9, (pp. 226-231) Cambridge Ma.,The MIT Press.
Bonabeau, E., & Dessalles, J . L. (1997). Detection and emergence. Intel-
lectica,25, 89-94.
Bonabeau, E., Dessalles, J . L., & Grumbach, A. (1995a). Characterizing
emergent phenomena (1): A critical review. Rev. Int. Syst. 9, 327-346.
Bonabeau, E., Dessalles, J . L., & Grumbach, A. (1995b). Characterizing
emergent phenomena (2): A conceptual framework. Rev. Int. Syst. 9,
347-369.
Boudon, R. (2006). Are we doomed to see the homo sociologicus as a ra-
tional or as an irrational idiot? In J . Elster, O. Gjelsvik, A. Hylland, &
K. Moene (Eds.), Understanding choice, explaining behavior. Essays
in the Honour of O. J. Skog (pp. 25-42) Oslo, Unipub Forlag, Oslo:
Academic Press.
Boudon, R. (1998). Social mechanisms without black boxes. In Hedstrm and
Swedberg (Eds.), Social mechanisms: An analytical approach to social
theory (pp. 172-203). Cambridge, MA: Cambridge University Press.
Bourgine, P., (1993). Models of autonomous agents and of their co-evolu-
tionary interactions; Entretiens J acques Cartier, Lyon.
Bunge, M. (2004). Clarifying some misundetstandings about social systems
and their mechanisms. Philosophy of the Social Sciences, 34(2), 371-
281.
Bunge, M. (1979). A system Concept of Society: Beyond Individualism and
Holism, Theory and Decision, 10(1/4) J anuary 13-30.
Bunge, M. (1977a). Emergence and the mind. Neuroscience 2, 501-509
Bunge, M. (1977b). Treatise on basic philosophy. III: Ontology: The furniture
of the world. Dordrecht: Reidel.
Broad, C. D. (1925). The mind and its place in nature. London: Routledge
& Kegan Paul
Campbell, D. T. (1974). Downward causation in hierarchically organized bio-
logical systems. In Ayala & Dobzhansky (Eds.), Studies in the philosophy
of biology: Reduction and related problems. London: Macmillan.
Cariani, P. (1991a). Adaptivity and emergence in organisms and devices,
World Futures 31, Gordon & Breach Science Publishers, 49-70.
Cariani, P. (1991b). Emergence and artifcial life. In Langton, Taylor, Farmer,
& Rasmussen (Eds.), Artifcial Life II. (pp. 775-797) Santa Fe Institute
Studies in the Sciences of Complexity Proceedings Vol. X, Redwood
City, Ca., Addison-Wesley.
Castelfranchi, C. (2000). Through the agents minds: Cognitive mediators
of social action. Mind & Society, 1(1), 109-140, March.
Castelfranchi, C. (1998a). Simulating with cognitive agents: The importance
of cognitive emergence. In R. Conte, J . Sichman, & N. Gilbert (Eds.),
Multi-Agent Systems and Agent-Based Simulation, Lecture Notes in
Computer Science (Vol. 1534, pp. 26-44), Berlin, Springer.
Castelfranchi, C. (1998b). Emergence and cognition: Towards a synthetic
paradigm in AI and cognitive science. In H. Coelho (Ed.), Proceed-
ings of the 6
th
Ibero-AmericanConferenceonAI:ProgressinArtifcial
Intelligence, Lecture Notes in Computer Science 1484, (pp. 13-26),
London, Springer.
Clark, A. (1996). Happy coupling: Emergence and explanatory interlock.
In M. Boden (Ed.), The philosophy of artifcial life. Oxford: Oxford
University Press.
Clayton, P., & Davies, P. (2006). The re-emergence of emergence. Oxford:
Oxford University Press.
Coleman, J . S. (1990). Foundations of social theory. Cambridge, MA: Har-
vard University Press.
Conte, R. (1999). Social Intelligence Among Autonomous Agents, Compu-
tational, and Mathematical Organization Theory, 5, 202-228.
Conte, R., & Castelfranchi, C. (1995). Cognitive and social action. London:
UCL Press.
Crutchfeld. (1994). Is anything ever new? Considering emergence. In Cowan
& Melzner (Eds.), Integrative themes. Volume XIX of Santa Fe Institute
Studies in the Sciences of Complexity, Reading, Ma., Addison-Wesley
Publishing Company.
Davidson, D. (1970). Mental events. In L. Foster & J . W. Swanson (Eds.),
Experience and theory. London: Duckworth.
Davidson, D. (1980). Actions and events. Oxford: Clarendon.
Deguet, J ., Demazeau, Y., & Magnin, L. (2006). Elements about the emer-
gence issue a survey of emergence defnitions. To appear in ComPlexUs,
International Journal on Modelling in Systems Biology, Social, Cognitive
and Information Sciences, Karger, Basel, Spring 2006.
Dennett, D. C. (1996). Kinds of minds. NY: Brockman.
Dessalles, J . L., & Phan, D. (2005). Emergence in multi-agent systems:
Cognitive hierarchy, detection, and complexity reduction part I: Meth-
odological issues. In P. Mathieu, B. Beaufls, & O. Brandouy (Eds.),
Agent-based methods in fnance, game theory, and their applications,
(pp. 147-159), Lecture Notes in Economics and Mathematical Systems,
Vol. 564, Berlin, Springer.
Dessalles, J . L., Galam, S., & Phan, D. (2006). Emergence in multi-agent
systems part II: Axtell, Epstein and Youngs revisited, 12th International
Conference on Computing in Economics and Finance,,J une 22-25, 2006,
Limassol, Cyprus
Drogoul, A., & Ferber, J . (1991). A behavioral simulation model for the
study of emergent social structures. EuropeanConferenceonArtifcial
Life, Paris.
Drogoul, A., Ferber, J ., & Cambier C. (1994). Multi-agent simulation as
a tool for analysing emergent processes in societies. Proceedings of
Simulating Societies Symposium, University of Surrey, Guildford, 1992.
Re-published in N. Gilbert & J . Doran (Eds.), Simulating societies:
The computer simulation of social phenomena, (pp. 127-142) Londres:
UCL Press.
Durkeim, E. (1895). Les rgles de la mthode sociologique, trad. by W.D.
Halls. The rules of the sociological method. New York: The Free Press,
1982.
Durlauf, S. N. (2001). A framework for the study of individual behavior and
social interactions. SociologicalMethodology,31(1), 47-87, J anuary.
Durlauf, S. N. (1997). Statistical mechanics approaches to socioeconomic
behavior. In Arthur, Durlauf, & Lane (Eds.), The economy as an evolv-
ing complex system II, Santa Fe Institute Studies in the Sciences of
Complexity, Volume XXVII, Redwood City, CA, Addison-Wesley.
Elster, J . (1989). Nuts and bolts for the social sciences. Cambridge MA:
Cambridge University Press.
Elster, J . (1998). A plea for mechanism. In P. Hedstrom & R. Swedberg (Eds.),
Social mechanisms. An analytical approach to social theory (pp. 45-73).
Cambridge, MA: Cambridge University Press.
Emmeche, C., Koppe, S., & Stjernfelt, F. (1997). Explaining emergence-
towards an ontology of levels. Journal for General Philosophy of Sci-
ence, 28, 83-119.
Epstein, J . M. (2006). Generative social science: Studies in agent-based
computational modeling. Princeton University Press.
Epstein, J . M. (1999). Agent-based computational models and generative
social science. Complexity,4(5), 41-60.
Ferber, J . (2007a). Towards an integrative view of multi-agent systems.
Forthcoming.
Ferber, J . (2007b). Multi-agent concepts and methodologies. In F. Amblard &
D. Phan (Eds.), Multi-agent models and simulation for social and human
sciences. Oxford: The Bardwell-Press (forthcoming in September).
Ferber, J . (1999). Multi-agent systems. Reading, MA: Addison Wesley.
Ferber, J ., Labbani, O., Mller, J . P., & Bourjault, A. (1997). Formalizing
emergent collective behaviors: Preliminary report. International Work-
shop on Decentralized Intelligent and Multi-Agent Systems (pp. 113-
122), St. Petersburg.
Forrest, S. (1990). Emergent computation: Self-organizing, collective, and
cooperative phenomena in natural and artifcial computing networks.
Introduction to the Proceedings of the 9
th
Annual CNLS Conference in
Emergent Computation (pp. 1-11). Cambridge, MA: MIT Press.
Galam, S. (2004). Spontaneous symmetry breaking and the transition to
disorder in physics. In P. Bourgine & J . P. Nadal (Eds.), Cognitive eco-
nomics, an interdisciplinary approach (pp. 157-168). Berlin-Heidelberg:
Springer-Verlag.
Galam, S., Gefen, Y., & Shapir, Y. (1982). Sociophysics: A mean behavior
model for the process of strike. Mathematical Journal of Sociology, 9,
1-13.
Gilbert, N. (2007). Computational social science: Agent-based social simula-
tion. In F. Amblard & D. Phan (Eds.), Multi-agent models and simulation
for social and human sciences. Oxford: The Bardwell-Press (forthcom-
ing in September).
Gilbert, N. (2002). Varieties of emergence. Paper presented at the Agent
2002 Conference: Social Agents: Ecology, Exchange, and Evolution,
Chicago.
Gilbert, N. (1995). Emergence in social simulations. In N. Gilbert & R.
Conte. op.cit. p. 114-156.
Gilbert, N., & Conte R. (1995). Artifcial societies: The computer simulation
of social life. London: UCL Press.
Gillett, C. (2002a). The varieties of emergence: Their purposes, obligations,
and importance. Grazer Philosophische Studien, 65, 89-115.
Gillett, C. (2002b). Strong emergence as a defense of non-reductive physical-
ism: A physicalist metaphysics for downward determination. Prin-
cipia, 6(1) J une, special issue on Emergence and Downward Causation,
83-114.
Granovetter, M. S. (1985). Economic action and social structure: The problem
of embeddedness. American Journal of Sociology, 91, 481-510.
Holland, J . H. (1998). Emergence from chaos to order. Readings, MA: Helix
Books Addison-Wesley.
Hollis M. (1994). The philosophy of social science. Cambridge, MA: Cam-
bridge University Press, (revised and updated version 2002).
Kampis, G. (1991a). Self-modifying systems in biology and cognitive science.
Pergamon Press.
Kampis, G. (1991b). Emergent computations, life, and cognition. World
Futures 31, Gordon & Breach Science Publishers, p. 33-48.
Kauffman, S. (1990). The sciences of complexity and origin of order. PSA:
Proceedings of the Biennal Meeting of the Philosophy of Sciences As-
sociation, Vol II Symposia and Invited Papers (pp. 299-322).
Kim, J . (2006). Emergence: Core ideas and issues. Synthese, 151(3), 347-
354.
Kim, J . (1999). Making sense of emergence. Philosophical Studies, 95, 3-
36.
Kim, J . (1995). Emergent properties. In Honderich (Ed.), The Oxford Com-
panion to Philosophy (pp. 224). Oxford University Press.
Kim, J . (1993). Supervenience and mind. Cambridge, MA: Cambridge Uni-
versity Press.
Kim, J . (1992). Downward causation. In Beckermann, Flohr & Kim, Eds.,
op cit.. 119-138
Kincaid, H. (1996). Philosophical foundations of the social sciences. Cam-
bridge, MA: Cambridge University Press.
Kincaid, H. (1986). Reduction, explanation, and individualism. Philosophy
ofScience,53(4) December, 492-513.
Kistler, M. (2006). New Perspectives on Reduction and Emergence in Phys-
ics, Biology and Psychology. Special issue of Synthese, 151/3.
Klee, R. L. (1984). Micro-determinism and concepts of emergence. Philoso-
phy of Science, 51(1), 44-63.
Labbani, O., Mller, J . P., & Bourjault, A. (1996). Describing collective
behaviors in Workshop ICMAS, 09-13 December, Kyoto, J apan
Langton, C. (1989). Computations at the edge of chaos: Phase transitions
and emergent computation. PhysicaD,42, p. 12-37
Langton, C. G. (1984). Self-reproduction on a cellular automaton. Physica
D,10, 135-144.
Li, M., & Vitanyi, P. M. B. (1993). An introduction to Kolmogorov complexity
and its applications. New York, Berlin: Springer-Verlag.
Lewes, G. H. (1875). Problems of life and mind. Rinehart & Winston.
Leyton, M. (2001). A generative theory of shape. New York: Springer Ver-
lag.
Manzo, G. (2007). Progrs et urgence de la modlisation en sociologie, du
concept de modle gnrateur et sa mise en oeuvre, LAnne Soci-
ologique, 2007, 57(1) forthcoming
McLaughlin, B. P. (1997). Emergence and supervenience. Intellectica,25,
25-43.
McLaughlin, B. P. (1992). The rise and fall of British emergentism. in Beck-
ermann, Flohr & Kim, op cit.. 49-93.
McLaughlin, B., & Bennett, K. (2006). Supervenience. In E. N. Zalta (Ed.),
The Stanford Encyclopedia of Philosophy (Fall 2006 Edition). Retrieved
from http://plato.stanford.edu/archives/fall2006/entries/supervenience
Mill, J . S. (1843). System of logic. London: Longmans, Green, Reader, and
Dyer.
Minsky M. L. (1965). Matter, mind, and models. Proceedings of IFIP Con-
gress (pp. 45-49), Spartan Books, Wash. D.C. Reprinted in Semantic
Information Processing MIT press, 1969.
Morgan, C. L. (1923). Emergent evolution. London: Williams & Norgate.
Morgan, M. S., & Morrison, M. (1999). Models as mediators. Cambridge,
MA: Cambridge University Press.
Mller, J . P. (2004). Emergence of collective behavior and problem solv-
ing. Engineering Societies in the Agents World IV, The 4
th
International
Workshop ESAW-2003, Revised Selected and Invited Papers, (pp. 1-20)
LNAI 3071 Springer Verlag.
Nagel, E. (1961). The structure of science. New York: Harcourt, Brace, and
World.
OConnor, T., & Hong Y. W. (2006). Emergent properties, E.N. Zalta (Ed.),The
Stanford Encyclopedia of Philosophy (Winter 2006 edition), http://plato.
stanford.edu/archives/win2006/entries/properties-emergent.
Palmer, R. (1989). Broken Ergodicity. In Lectures in the Sciences of
Complexity, D. Stein, Ed. Volume 1, Santa Fe Institute Studies in the
Sciences of Complexity, Reading, Ma., Addison-Wesley.
Phan, D. (2004a). From agent-based computational economics towards
cognitive economics. In P. Bourgine, & J . P. Nadal (Eds.), Cognitive
economics, An interdisciplinary approach (pp. 371-398). Berlin-Hei-
delberg: Springer-Verlag.
Phan, D. (2004b). Hierarchy of cognitive interactive agents and statistical
mechanics: How object oriented programming highlights the connec-
tion. In Coelho, Espinasse, & Seidel (Eds.), The 5
th
Workshop on Agent
Based Simulation, Lisbon, Portugal (pp. 69-76) Erlangen, San Diego,
SCS Pub. House.
Phan, D., & Feber, J. (2007). Thinking the social dimension of the artifcial
world: Ontological status of collective beliefs. International Transac-
tions on Systems Science and Applications, Special issue on Emergent
intelligence over networked agents, forthcoming. in 2007.
Phan, D., & Semeshenko, V. (2007). Equilibria in models of binary choice
with heterogeneous agents and social infuence. European Journal of
Economic and Social Systems, forthcoming.
Phan, D., Gordon, M. B., & Nadal, J . P. (2004). Social interactions in eco-
nomic theory: An insight from statistical mechanics. In P. Bourgine, &
J . P. Nadal (Eds.), Cognitive economics, An interdisciplinary approach
(pp. 225-358). Berlin-Heidelberg: Springer-Verlag.
Philemotte, C., & Bersini, H. (2005a). Intrinsic emergence boost adaptive
capacity. In GECCO 05: Proceedings of the 2005 Conference on
Genetic and Evolutionary Computation (pp. 559-560) New York, NY,
USA, ACM Press.
Philemotte, C., & Bersini, H. (2005b). Co-evolution of effective observers
and observed multi-agents system. AdvancesinArtifcialLife,The8
th

EuropeanConference,ECAL2005, Canterbury, UK, September, Pro-
ceedings. (pp. 785-794) Volume 3630 of Lecture Notes in Computer
Science, Springer.
Poincar, H. (1902). La science et lhypothse, Paris, Flamarion. Trad. Sci-
ence and Hypothesis, New York, Dover Publications 1952
Putnam, H. (1981). Reason, truth, and history. Cambridge, MA: Cambridge
University Press.
Ronald, E., & Sipper, M. (1999). Surprise versus unsurprised: Implication of
emergence in robotics. RoboticsandAutonomousSystems,37, 19-24.
Ronald, E., Sipper, M., & Capcarrre, M. S. (2001). Design, observation,
surprise! A test of Emergence. ArtifcialLife,5, 225-239.
Rosen, R. (1985). Anticipatory systems: Philosophical, mathematical, and
methodological foundations. New York: Pergamon Press.
Rosen, R. (1978). Fundamentals of measurement and representation of
natural systems. North Holland.
Rosen, R. (1977). Complexity as a system property. International Journal
ofGeneralSystems,3, 227-232.
Rueger, A. (2000a). Physical emergence, diachronic, and synchronic. Syn-
these,124, 297-322.
Rueger, A. (2000b). Robust supervenience and emergence. Philosophy of
Science,67, 466-489.
Sawyer, R. K. (2005). Social emergence: Societies as complex systems.
Sawyer, R. K. (2004). The mechanisms of emergence. Philosophy of the
Social Sciences, 34(2), 260-282, J une.
Sawyer, R. K. (2003). Nonreductive individualism, Part II: Social causation.
PhilosophyoftheSocialSciences,33(2), 203-224
Sawyer, R. K. (2002a). Emergence in psychology: Lessons from the history
of non-reductionist science. HumanDevelopment,45(2), 1-28.
Sawyer, R. K. (2002b). Nonreductive individualism, Part I: Supervenience and
wild disjunction. PhilosophyoftheSocialSciences,32(4), 537-559.
Sawyer, R. K. (2001a). Emergence in sociology: Contemporary philosophy of
mind and some implications for sociological theory. American Journal
ofSociology,107(3), 551-585.
Sawyer, R. K. (2001b). Simulating emergence and downward causation in
small groups. In Moss & Davidsson (Eds.), Multi agent based simula-
tion (pp. 49-67). Berlin, Springer-Verlag.
Schelling, T. S. (1978). Micromotives and macrobehavior. NY: W.W. Norton
and Co.
Schelling, T. S. (1969). Models of segregation. American Economic Review,
Papers and Proceedings, 59(2), 488-493.
Simon, H. A. (1996). The science of the artifcial. Cambridge, MA: MIT
Press.
Sperry, R. W. (1991). In defense of mentalism and emergent interaction.
JournalofMindandBehavior,12, 221-245
Sperry, R. W. (1986). Discussion: Macro- versus micro-determinism (A
response to Klee). PhilosophyofScience,53(2), 265-270.
Sperry, R. W. (1969). A modifed concept of consciousness. Psychological
Review,76, 532-536
Steels, L. (2006). Experiments on the emergence of human communication.
TrendsinCogn.Sciences,10(8) 347-349.
Stephan, A. (2002a). Emergentism, irreducibility, and downward causation.
Grazer Philosophiche Studien, 65, 77-93
Stephan, A. (2002b). Emergence. In Nadel (Ed.), Encyclopedia of Cognitive
Science (Vol. 1, (pp. 1108-1115). London: Macmillan.
Teller, P. (1992). A contemporary look at emergence. In Beckermann, Flohr,
& Kim, Eds., op cit. 139-153
Tesfatsion, L. (2002a). Economic agents and markets as emergent phenomena.
Proceedings of the National Academy of Sciences U.S.A., 99, Suppl. 3,
7191-7192
Tesfatsion L. (2002b). Agent-based computational economics: growing
economies from the bottom up. ArtifcialLife,8(1), 55-82.
Tesfatsion, L., & J udd, K. L. (2006). Handbook of computational econom-
ics.Vol.2:Agent-BasedComputationalEconomics. Amsterdam, New
York: Elsevier North-Holland.
Tuomela, R. Methodological Individualism and Explanation, Philosophy of
Science, 57, 133-140.
Udehn, L. (2001). Methodological individualism: Background history and
meaning. New York: Routledge.
Van de Vijver, G. (1997). Emergence et explication. Intellectica,25, 7-23.
Varela, F., Thompson, E., & Rosch E. (1991). The embodied mind. MIT
Press.
Vaario, A, Hori, K., & Ohsuga, S. (1995). Towards evolutionary design of
autonomous systems. The International Journal in Computer Simula-
tion, 5, 187-206.
Wilbert, K. (2000). A theory of everything: An integral vision for business,
politics, science, and spirituality. Boston, MA: Shambhala Publica-
tions, Inc.
Wolfram, S. (1984). Universality and Complexity in Cellular Automata ,
Physica D. N10.
Worrall, J . (1989). Structural realism: The best of both worlds? Dialectica,
43, 99.
Zahale, J. (2003). The individualism-holism debate on intertheoretic reduction
and the argument from multiple realization. Philosophy of the Social
Science,33(77), 99.
Endnotes
1
For OConnor and Hong (2006), the contemporary discussion on
epistemological emergence as a non-reductionist concept relies on the
discussion on Nagel (1961, p. 366-380 The doctrine of emergence).
The latter is in fact widely based on the British version of emergentism,
with Broad and Mill.
2
For mathematical Platonicists, mathematical entities exist independently
of the human observer. The strong form of scientifc realist is also
called metaphysical realism by Putnam (1981). This reject does not
imply a relativist position. Our position is compatible for instance with
the internal realism of Putnam (1981) as well as with more anti-rela-
tivist positions such as the structural realism of Poincar (1902) and
Worrall (1989)see Varenne, Phan epistemology in a nutshell in
Amblard & Phan, (2007).
3
See Fromm, in this book, for another typology of emergent phenom-
enon.
4
For a recent debate including the supervenience/reduction dimension
discussed here, see Kinkaid (1986) and Tuomela (1990).
5
Energy is a resultant or hereditary property in the sense that it is pos-
sessed by every parts of a thing (...) Unlike holism, we regard emergents
as rooted in the properties of the components, hence as explainable in
terms of the latter, through not by reduction to them. For example, tem-
perature is explainable as average kinetic energy of the molecules, but
this does not exemplify reduction because averages, through computed
from individual values alone, are collective properties. (Bunge, 1977b
p. 97-98)
6
For a discussion of both perspectives in the social sciences from a
median emergentist point of view, see the antagonists contributions
of Archer (1998) and Sawyer (2001a)
7
See for instance Durlauf (1997, 2001), Phan, Gordon, and Nadal (2004),
Phan and Semeshenko (2007), and the pioneering work of Galam, Gefen,
and Shapir (1982).
8
Remark that several authors consider irreducibility as a necessary
condition of emergence. Accordingly, numerous phenomena studied
by the statistical mechanics are not viewed as emergent. For instance,
for Bunge (1977), the temperature does not emergent from molecular
movements because it could be reducible to the average energy within
the system.
9
An important feature is the availability of the inferior level of cognition
for higher-level agents: an epistemic agent can behave sometimes like
a behavioral agent or like a reactive agent.
00 Alker
Chapter.X
Ontological Refections on
Peace and War
1
Hayward R. Alker, Unversty of Southern Calforna &
Watson Insttute, Brown Unversty, USA
Abstract
Responding to a provocative question by Hiroharu Seki about Hiroshima
ontologies, this chapter reviews related thinking about the ontological
primitives appropriate for event-data making, accessing high-performance
knowledge bases, and modeling intelligent complex adaptive systems of use
to researchers on war and peace. It cautions against Cliocide, defned
as of the silencing or symbolic killing of collective historical-political
or historical-disciplinary identities and identifying practices by historical
ordisciplinedefcientscientifccodingpractices.Itproposesthatmore
intelligent multi-agent models in the complex, adaptive systems tradition
of the Santa Fe Institute should include the socially shared memories of na-
tionsandinternationalsocieties,includingtheiridentity-redefningtraumas
Ontological Refections on Peace and War 301
and their relational/migrational/ecological histories of community-building
success and failure. Historicity in an ontologically distinctive sense of the
time ordered self-understandings of a continuing human society is still a
challenge for the computationally oriented literature on war and peace.
Introduction.

Faced with the challenge to honor the late Hiroharu Seki, whom I knew and
have admired for several decades, I was at frst not sure what would be an
appropriate appreciation of his signifcant, wide-ranging career. The theme
of ontologies relevant to researchers interested in peace and war fnally
crystalized in my mind followed hearing him speak from the audience at a
recent International Studies Association meeting. Politely, but insistently,
he asked a distinguished panel of scholars (I paraphrase): how can you put
Hiroshima ontology into your approach to international relations? At a panel
where epistemological and ontological pluralism was a central issue, and
before an inter-disciplinary professional group where socially constructed
international relationships have been a hot topic
2
, his remark was both highly
appropriate and particularly challenging.
Sekis diffcult but powerful insistence deserves sustained refection. I see
his injunction as closely related to the development of a collective sense
of history as a basis for both an international society of states and a global
society of individuals having the capacity to direct themselves towards a
better future, away from horrifying pasts, toward a world where Hiroshimas,
Auschwitzs, and Chernobyls never reoccur. In other words, Seki appeared
to be trying to discover how can a powerful, shared sense of history and
tradition, of prohibitions and morally preferable possibilities, be developed,
shared, and deeply transmitted to future generations. I think he had in mind
their culturally transmitted, phenomenological, and ontological ways of see-
ing and encoding reality, as well as their more consciously refected upon
norms and practices.
To pay homage to Professor Seki, I have decided to make some earlier
thoughts along similar lines more widely available. First, I shall recall briefy
the international relations simulation world of the 1960s, which Hiroharu
0 Alker
Seki and I both inhabited, and where we frst met through the courtesy of
Harold Guetzkow. Then, I shall refect further on the absence of morally
powerful lessons easily connected to, or embodied within, the way behav-
ioral scientists have represented specifc, singular international events like
Hiroshima or Auschwitz. How is this kind of intelligence embodied into our
complex adaptive systems models? Where, for example, in formally modeled
accounts of international relations, are the triggers for the tragic (or comic!)
meanings of many classical historians, or the ironical historical sensitivity
characteristic of many modern and some post-modern thinkers?
3
And what
about the data we develop to calibrate or test such models? Do not simple
quantifcations of horrifying international events like these obfuscate, or tend
to erase from our consciousness, the complicated socio-historical identifca-
tions, expectations, roles and networks normally binding together a national
or international societys members?
Current adaptive systems models popular in political science and international
relations have no or almost no symbolic memories; typically their past-in-
voking mechanisms are structurally similar to frst order Markov processes
4
.
Surely, the human history of peace and war, of freedom and domination, of
community, nation, state or international systems formation and dissolution
has more to offer than that! I shall suggest event data coding practices
5
for
greatly reducing the inadvertent complicity with such horrible events, which
their oversimplifed, reality hiding, naturalizing scientifc codifcations,/
quantifcations as event data tend to suggest. I shall follow the practice of
contemporary computer scientists in treating ontologies for such event data
as humanly respecifable linguistic primitives used for artifcial knowledge
representation and retrieval systems.
6
Next, attempting to help fulfll Profes-
sor Sekis belief in the promise of international simulations, I shall refect on
my several visits to a citadel of complex adaptive systems theory, the Santa
Fe Institute. I shall show how a better sense of historical connectivities, or of
trauma-avoiding historicities (a concept to be defned below), might inform
a recent ecological research program on adaptive, multi-agent systems,
7
a
research program attempting to incorporate international issues. Finally, as
a challenge to future modelers of intelligent complex adaptive systems, I
speculate about the future of more intelligently modeled international crisis
research, inspired by the recent computationally-oriented writings of Gavan
Duffy, Brian Cantwell Smith, and Thomas Schmalberger.
Putting More War and Peace into
Our.Models.Ontologies
In the late 1960s,when Professor Seki and I were both associated with
Harold Guetzkows simulated international processes projectI made my
frst investigation of phenomenological and ontological possibilities in more
or less computerized international simulations.
8
In what then I might have
referred to as a study of the face validity of different simulated worlds,
my investigative strategy was comparatively to attempt to represent the
consequences of an accidental, rather than an intentional, nuclear explosion
in three different simulated worlds.
But even a hypothetical, gamed exploration of the implications of an ambigu-
ously caused
9
Cold War nuclear accident at a clandestine Soviet test site in
Polunochnoe, proved beyond the representational capacities of skilled formal
modelers at that time. Professor Lincoln Bloomfelds non-computerized Politi-
cal Military Exercise evidenced much more realistically suggestive possibilities
for escalation and creative confict containment than the formalized or partly
formalized models Ronald Brunner and I investigated. The crucial gamed
results of Bloomfelds imaginative scenario generate diffculties in the U.S.
governments communication, for verifcation purposes during a crisis situa-
tion, with submerged submarines. These diffculties could not be adequately
represented within either Guetzkows own Inter-Nation Simulation (INS) or
a hugh, Defense Departments TEMPER computer simulation. Guetzkows
man-machine INS combined human role players with computational de-
cision environments, abstracted actors, and relationships too far above the
important particulars of international history. It treated nations as abstract
entities named Bingo, Erga, etc. Similarly, the major fully computerized
simulation at the U.S. Department of Defenses J oint War Games Agency in
the late 1960s, TEMPER, couldnt represent unintentional actions well; it
also regionalized international threats implausibly. Fulflling treaty obliga-
tions, European NATO members sent troop reinforcements to the continental
United States in the middle of the TEMPER-simulated Polunochnoya crisis!
Sekis ontological challenge was already a relevant issue, although the devel-
opers of TEMPER thought historically in categories linked most directly to
the Cold War rather than to the costs of victory and defeat in World War II.
A look at event data research in subsequent decades would reveal, I believe,
persistent versions of related representational conundrums.
10

0 Alker
How.not.to.Commit.Cliocide.in.
Event-Data.Coding.Practices
11
.
As methodologically restated in Alkers Emancipatory Empiricism: Toward
the renewal of empirical peace research,
12
yesterdays computational tools
and data-making practices rarely have the avoidance of specifc major inter-
national traumas (which actually occurred much earlier and might recur today
or in the future) as their central, codifable reality. Our realistic scientifc
procedures have long been in danger of being captured by the often ugly,
frequently recurring but probably not eternal realities of Power Politics that
we as scholars of the successes and failures of the Westphalian state system
must try faithfully to represent, as well as to transcend.
What are the events that the event-data specialist tries replicably, and
hopefully validly, to encode? Websters Third New International Dictionary
illustrates its frst defnition of event as something that happens by citing
Wordsworththis days event has laid on me the duty of opening my heart.
Refect on his indictmentin a poem entitled The Thornof science (and
art) unrefective about the infuences of its originating experiences:
Our meddling intellect
Misshapes the beauteous forms of things.
We murder to dissect.
Enough of science and of art.
Close up these barren leaves.
Come forth and bring with you
A heart that watches and receives.
Whether beauty or ugliness is our subject, the naturalistic search for quantifed
scientifc replicability can butcher specifcally situated traumas or ecstacies.
We must learn from Wordsworth.
Two themes derive from these refections on the ways in which we fail
scientifcally to encode the partly shared, socio-historical signifcances of
Hiroshima, Auschwitz, and other traumatic events shaping the internal and
relational identities and international destinies of particular nations or
peoples. First, I will briefy discuss the high politics of disciplinary birth,
renewal, death and transfguration. In this domain paradigmatic, disciplin-
ary, and interdisciplinary wars have sometimes been directed more or less
consciously against humanistic history, radical political economy, transna-
tional/international sociology, peace research, the cultural self-understand-
ings of conquered nations, and even the social reality of power politics.
None of these orientations adequately comprehends the complexities of
international history.
The second theme concerns the ways International Relations special-
ists operationally address event data. It concerns how they might
better include both socio-cultural and political-economic historicities,
including the recognition of distortive effects of their own disciplinary
historicity, in their work. This new, merged research direction, now sig-
nifcantly, but never completely, operationalizable with contemporary
multi-perspective procedures
13
, should help slow or reverse our disci-
plinary descent from analytical accident to malfeasance to symbolic
silencing or Cliocide. By historical/disciplinary Cliocide, I mean the
more or less premeditated analytical and/or symbolic killing of socio-his-
torically shared achievements and failures, including the replicable traumas
and the people-defning, semi-historical hero stories of societal collectivities
and transnational disciplines. The culprits?: solipsistic, individualistic, one-
sided, asocial, scientistic, phenomenologically insensitive and ontologically
impoverished story-telling and data-making perspectives and procedures.
The.Historicity.of.International.Events.and.Scholarship.

According to Webster, then, events are things that happen, activities or experi-
ences, noteworthy occurrences or happenings; they are thus pieces of history.
International events are then events that are international in some respect,
pieces of international history noteworthy from perhaps several domestic or
international perspectives (usually those of journalists). International event
data are replicable codifcations of these pieces of historical experience,
generated from particular international relations scholarly vantage points.
But, we must ask, what are the boundaries of an event, what are the con-
nections (at any instant, and through time), and how do they affect the kind
of event it is, or that we, or others, describe it to be? How can we better
delineate the ways events contain, overlap, inter-relate, interpenetrate, fold
back upon, cumulate, or subsume each other? Our scholarly impoverishment
in ontologically conceiving or phenomenologically representing historys
essence, its historicity (or, more generally, its connectivity) is evident in
0 Alker
the small number of ways most complex adaptive systems modelers of war
and peace conceive and represent historical international processes. And by
the extent to which historicity is usually naturalistically defned as a kind of
path dependence based on the model of sampling from urns of marbles
without replacement.
Let us look at the issue of historical connectivity, or historicity, in somewhat
more detail. Some refer to the Pelopponesian War as an event; similarly others
refer to World War I or II separately, the latter being argued to have started in
1931 (Manchuria), 1935 (Ethiopia), 1939 (Poland), or 1941 (Pearl Harbor).
But classical historians recognize that, just as world systems theorists are
having us rethink the ultimate separability of the First and Second World Wars
of this century, Thucydides was both insightful and correct to treat as one war
two periods of sustained hostilities interrupted by a long, but uneasy peace.
The meaning and signifcance of the later events in both warsin particular,
as attempts to fulfll, redefne, or overturn earlier policies or settlements (i.e.,
positively or negatively judged precedents)is so colored by the earlier pe-
riod of fghting that it is appropriate to say the earlier events are contained in,
or constitutive of the later ones. Together these episodes constitute a single
event, a larger unity. Here is a conception of historicitythe time ordered
self-understanding of a continuing human society
14
that further and more
deeply explicates Sekis diffcult conception of Hiroshima ontology.
We can summarize and extend the present discussion with Table 1. Reaching
back into recent philosophical and historiographic writings, and toward the
technically advanced felds of (computational) linguistics gives at least nine
overlapping, but alternative versions of historicity.
One may well ask, must any historical particular always be defned in terms
of everything it is connected to? Which perspectives concerning an eventat
least including signifcant contemporary participantsdo not belong in its
adequate codifcation? Behavioral codings by one professors graduate students
debatably assume (and allow limited statistical tests for) conceptual univer-
salism. But the diversity, multiplicity and richness of multiple perspectives
concerning war boundaries is already powerfully suggested, inter alia, by
the continentally distinct initial dates for World War II previously given. As
recent American history, and the diplomatic Bluebooks of World War I fame
make clear, sharply different versions of events, from different individual,
organizational, national or transnational perspectives, are intriguingly, but
differentially constitutive of their noteworthiness. Overlapping, changing, as
well as opposed, regionalisms
15
seem a better empirical starting point for the
Table 1. Nine ways of conceptualizing and/or operationalizing historicity
(Source:Mostimportantly,Alker,1996,p.184-206,213-215,297-302,321-
328,and386-393)
1. The Brian Arthur-Paul David-Robert Keohane-Stephen Krasner literature
on cybernetic, path-dependent, probabilistic (urn selection) historical
processes;
2. Weiner-inspired cybernetic discussions of nonergodic development;
3. Jon Elsters rewrites of electro-magnetic hysteresis (defned by Webster
as the infuence of the previous history or treatment of a body on its
subsequent response to a given force or changed condition).
4. The philosophical-political traditions of practical argumentation begun
by Thucydides pre-Socratic teachers and distilled in recent technical-
philosophical literatures by Nicholas Rescher, Ronald Loui and Douglas
Walton on practical and/or dialectical argumentation;
5. Derivation-dependent parsing, already clearly illustrated in Chomskys
Syntactic Structures and also visible in the context-sensitive rewrite rules
of text linguistics;
6. The artifcial intelligence literature on non-monotonic reasoning: non-
monotonic reasoning has the historically suggestive properties: theses are
tentative & reversible; axiomatic presumptions may change in the course
of arguments; thus previously derived conclusions can be invalidated;
7. The self-compiling and self-modifying capacities of computer procedures
or programs written in the lexically scoped SCHEME dialect of LISP,
programs which have the additional property of transparent historicity,
process traceability back to original defnitions and contexts of defni-
tion;
8. The ethnomethodological/phenomenological literature on scientifc
development and members time;
9. Hegels phenomenologically inspired discussions of Reason and Spirit
in History, Vicos and Marxs variants of historicism, and Heidigger and
Marcuses ontological discussions of historicity, from which Olafsons
previously cited, analytically much clearer conception of historicity partly
derives.
0 Alker
boundaries, parties, and issues making up some of the more complex wars
of the present (and other) centuries. We need critically and constructively to
rethink the bases for excluding from critical, comparative, scientifc analyses
ab initio certain nationalities, traditions or paradigms of research in the dif-
fcult, multi-disciplinary feld of international relations research.
Part of the historical signifcance or noteworthiness of international events
arises from the larger events, processes, histories or which they are seen to
be part. Contemporaries can give signifcance to the events of their times in
terms of larger (but debatable) historical constructsThucydides account
of Athens imperial rise and tragic fall, Aristophanes proto-feminist indict-
ments of the destructiveness of male Athenian economic greed come to mind.
Historicist Soviet versions of the triumph of Socialism are now obviously bad
history compared to those interpretations of specifc recent events consistent
with the large theme of the rise of global capitalism. But at what point, and
why, should such internal, identity-constitutive connections to larger, debat-
able historical happenings be ignored?
In our scientifc codifcations of signifcance, we must remember that distances
and historical projects are measured differently, with different degrees of
universality and timelessness. From a multi-disciplinary perspective, is not
the genre of classical authors writings indicative of a multi-faceted approach
to truths of various sorts? Aristophanes entered his plays into the Athenian
competition, presumably describing them as comic fctions; but they serve
as evidence for feminists and Marxists of economic and gender differences
neglected by Realist analyses. And although Thucydides scientifc his-
tory obviously can be distinguished from crowd-pleasing morality plays,
that doesnt destroy Raymond Arons claim that Thucydides wrote art and
science simultaneously, or the Classical reading of this history as a tragic
morality play about the arrogance of power.
16

What unbiased science can legitimately discard the evidence presented
by any of these perspectives? Ironic criticism has been a mainstay of pro-
gressive, critical, modern thought. Must we try to denude our language of
its morally based critical power? Supposedly in the interests of analytical
clarity, logical positivism tried to do so, but failed. Shouldnt a deeper ap-
preciation of how language works be a better guide? The particular kind of
event connectivity that most concerns us here was referred to above with
terms like socio-historical, historical sociality, or the historicity of
international events. Among philosophers of social science and historians,
historicity has been a concept developed to delineate both a very precious,
distinctive, and central feature of human socio-cultural evolution and of a set
of humanistic disciplines (history, rhetoric and dialectic, etc.) substantively
and methodologically distinctive from the natural sciences.
Admitting that further historical investigation of this argument is required,
I suggest tentatively that the relative lack of attention to the conceptual and
representational issues raised by this focus has several roots. It is linked both
to the logical positivist or behavioral scientifc origins of early events data
research methodologies, to a resistant unfamiliarity of most International Rela-
tions researchers with these professional-identity-challenging issues, and/or
to their lack of expert awareness of operationally relevant, but demanding,
alternative representational approaches like those in Table 1. More will be
said next about the paradoxical possibility that more technically sophisticated
representational/codifcation practicesthose described within a paradigm of
conversational computing
17
can give us more humanistically defensible
empirical research practices.
The historicity of events takes place through what phenomenologists refer
to as the intentional (in the double senses of object-constituting and pur-
posive), the extensional (referential) and the intensional (meaning constitu-
tive) dimensions of meaning, and the interpretive ways groups of citizens
and scholars recognize and defne the eventss existence. It takes place, one
might say metaphorically, when a social groups past is born (i.e., it be-
comes accessible to them in a temporal, but necessarily clock-like, fashion).
Following Olafson, as articulated in Table 1, we defned historicity as the
socially shared self-temporalization of their collective experiences among
members of a continuous human society of some kind. Harold Lasswell, for
example, clearly called for the empirical tracking of the demands, expectations
and identifcations that are part of these in such historicities. To summarize
his views, for better or worse we are embedded in historical confgurations
which are characterized by the existence of a large number of comprehensive
symbols in the name of which people kill or die.
18

A second class of nationally and transnationally shaped historicities should
also be mentioned: those of disciplinary societies of international relations
scholars, including the overlapping professional groupings of historians,
war researchers, and peace researchers. Among such inter-disciplines,
key symbol confgurations, research-orienting concepts and their changing
interpretations are central embodiments of professional historicities.
0 Alker
The connections of socio-political and disciplinary historicities are compli-
cated. Indeed, widely but incompletely shared concepts, paradigms of theory
and practice, and specialized symbols and languages refect and help constitute
national, transnational and international self-identifcations, understood both
historically and in disciplinary ways. To cite a Communitarian, a Realist and
a Radical/Marxist theorist, recall the notion of historicity implicit in Karl
Deutschs theory of pluralistic security communities, Hedley Bulls histori-
cal accounts of the expansion of the anarchical society around the globe,
or Lenins ideas about the revolutionary historicity of national Communist
parties and Socialist Internationals in the age of imperialism
19
.
When it is not practically or technically necessary to do so, to ignore the
collaborative and/or antagonistic historicities just referred to is defnitionally
or methodologically to bias, to distort, to silence the multiple meanings
imbedded within international historical reality. To codify or measure such
events in a correspondingly partial, unconnected fashion is to sever, cut off,
dismember or even quantitatively butcher essentially qualitative histories in
the good name of science. It is to destroy the transnational and international
historicities out of which international political community and world order
might eventually grow.
Operationalizing.Historicity.in.Event.Data.Research

Event data research attempts to apply scientifcally justifable and replicable
procedures to the categorical or quantitative measurement of relevant news-
paper accounts. It is a response, so far of limited but still urgent practical
utility, to the need to address the great majority of historical materials, which
are themselves qualitative in nature.
20
Having spent several decades research-
ing the issues of humanistic interpretation and explanation of the practice
and transcendence of Power Politics, I believe that we need still further to
develop much farther qualitative ways of modeling collective human inten-
tions, interpretations, and differentially shared precedential memories.
A practical variant of Sekis ontological question still remains: How can in-
ternational event data researchers be more deeply sensitive to the economic,
cultural and political historicities in terms of which social groups give meaning
to their collective experiences, in the name of which their typical members
are willing to kill, to die, or to be killed? The regulative ideal I suggest for
such work is develop computerized representations of socially signifcant
events that simultaneously contain or reference most or all of their external
and internal (identity constituting) relations, their inter- and inner-connec-
tions, and the cooperative and confictful historicities giving them social and
scholarly signifcance.
What are these relationships? Where can they be found? I have argued:
search through human memories and their documentary external stabilizers,
look at the constitutive conversations, the scholarly debates, the meaningful
interactions, the social discourses and even the negotiated silences where
the events become, remain, cease to be socially, culturally, economically
or politically signifcant, either historically or scientifcally. Because of the
way it emphasizes historic interconnections, and includes the history of its
generation, utilization and transformation, the relevant data set would look
like a well organized, and transparent human memory. Additional research
might try to fll in missing intentions, search for more holistic historical pat-
terns, paralleling the ways that our own passionate or relatively detached
intellects dissect and rewrite the past.
The very large role of natural language representations (which include and
can generate quantifcations where appropriate) is itself noteworthy. Data
analyses according to different paradigmatic codifcations and more richly
comparable than mere numbers. Historical narratives written from differ-
ent participant or scholarly perspectives provide levels of information and
modalities of interpretive coherence not naturally representable as statistical
coeffcients or dynamically changing, differently colored, grid boundaries.
When episode histories and codebooks can be treated as applicable precedents
or case-redefning procedures, conversationally refective modes of case- or
explanation-based reasoning
21
and data analysis can more fully parallel the
unsilenced conversational practices of emancipatory dialogical relations. The
linking of freedom, critical rationality and the responsible pursuit of larger
human goals goes back at least to Thucydides history.
This idealized sketch shows several ways in which the historical context of
particular events can be treated as part of that event, without the data ana-
lyst becoming, as Harold Guetzkow might say, data bound. Histories of
event citations in precedential scholar/practitioner texts or speeches would
augment existing data sets. Evidence on the various ways in which practical
historians or scholars construct their pasts would also grow. Evidence on
multiple historical and scholarly temporalizations would allow empirically
grounded, constructive characterizations of the shared historicity of social
events and social scientifc practice. Thus the sociality of shared, opposed,
Alker
or coordinated perspectives and practices would be social facts within the
data set, rather than statistically problematical realities.
Starting with work at MIT, and continuing at MIT and Syracuse, Gavan Duffy
and J ohn Mallery developed a RELATUS text analysis computer system
for computationally parsing, storing, or retrieving, and analyzing different
narrative accounts; they saw the need to extend these procedures to herme-
neutic, conversational and historical contexts of application,
22
which can be
organized in terms of different theoretical or socio-historical perspectives.
Measurements corresponding to codebook defnitions or revisions can be
reconstituted into the descriptive sentences from which they were initially
derived. Replicable, non-quantitative representations for catching elements
of the past in the present are beginning to appear. Historicity, potentially,
exists within a symbiotic network of human experts and their recursively
structured computer programs.
Incorporating.Political.Histories.into.Adaptively.
Intelligent.Multi-Agent.Ecological.Models
I have extended Sekis concern with the phenomenological and ontological
dimensions of a deep commitment to the avoidance of nuclear war destruc-
tion like Hiroshima to include genocides like that exemplifed by Auschwitz.
And I have argued that such concerns can be treated as specifc historicities.
Can this way of thinking be scientifcally extended to include the avoidance
of ecological devastation? Can such precedential historicities be included
at the foundational/ontological level in other research programs than those
International Relations has normally concerned itself with?
Coincident with the ages concern with better understanding and promoting
sustainable development, I now want to discuss ecologically oriented mod-
els, and the need to put sharable human histories, with their trauma-avoiding
lessons, into them. As Professor Seki would want us to do, I assert here, and
try to exemplify, the continuing need for interdisciplinary researchers on war
and peace to bridge gaps, to connect war and peace practices and possibili-
ties, to the issues of contemporary concern. As the previous pages should
have made apparent, the ontological and representational issues of putting
shared human consciousness, memories and moral choices within ecological
models is not a trivial one.
The need to historicize ecological modeling practices in a peace-relevant way
arose in my mind at an exciting April 1994 workshop which I attended at the
Santa Fe Institute in New Mexico (SFI).
23
There were at least three major
thematic convergences, pregnant with further research implications, evident
to me by the end of (or shortly after) my participation in the Crude Look at
the Whole (CLAW) Workshop. Here is a brief sketch of them.
Lateral.Pressure.Theory.Meets.Adaptive.Multi-Agent.
Modeling
The joint presentations of J ohn Holland and J oshua Epstein were a wonderful
kick-off for the seminar. As I soon learned, these were mere introductions to
a style of adaptive, multi-agent, grid-based, ecologically inspired modeling
that is a central thematic unity of the work at the SFI
24
. My background in
elementary artifcial life simulations and Hollands genetic algorithms meant
that I was not totally surprised by either presentation, but the interdisciplinary
synthesis of elements of geography, biology, demography, economics and
sociology in Epsteins stunningly effective visual presentation was very impres-
sive. The spatialization of social processes emergent from the interactions of
autonomous, message passing agents recalled writings by Schelling, Bremer
and Mihalka, Cusack and Stoll, and Duffy
25
on power political interactions
on territorial grids. Some of this literature was unfamiliar to SFI scholars.
And, driven by the concern for improving the peace relevance of the social
ontologies used in inter-disciplinary futures oriented work, I engaged with
these authors.
It was my impression that the Epstein-Axtell model had much simpler, but
more social scientifcally interesting bio-social foundations. But, I found
Hollands ECHO system (with its evocations of Burks, Axelrod, Cohen and
HollandMichigan BACH) elegant and visually suggestive, with a deeper
grounding in evolutionary ecology. The coloring of different Prisoners
Dilemma players on a spatial grid gave very nice cooperative results of
the emergent sort one has come to expect from SFI. Hollands and Murray
Gell-Manns
26
emphasis on the minimally complex cartoon-like character
of such formalizations/visualizations was entirely in the spirit of what I took
to be a heuristic exercise. Hollands recognition of the modular character of
genetic selection processes within selection processes recalled as well the
Leibnizian ontology (monads within monads...) informing the Rapoport/DT
Alker
Campbell/Boulding/Simon hierarchies of knowing/adapting processes I
learned from the General Systems Yearbook many years ago.
27

The initial convergence of Epsteins artifcial histories of life on his simulated
sugarscape with Choucri-North lateral pressure accounts of the demographic-
technological-resource-using growth, intersection and confict of nations was
stunning. Choucri and Norths econometric analyses of processes leading
up to World War I and their work on the imperialistic and nonimperialistic
ways that J apan expressed its lateral pressure before and after World War
II came immediately to mind.(Choucri & North, 1975; Choucri, North, &
Yamakage, 1992). And the demographic-technological-territorial profles in
(Choucri, 1993)
28
(which correspond roughly to different population densities
differently located on Epsteins sugarscapes) all ft Epstein-Axtells story
extremely well. The best evidence of a closeness of ft came when Robert
North and J oshua Epstein started talking about a common problem: how to
build hierarchical social structures into their expansionary stories. A kind
of class structure seemed to appear on Epsteins video screen, as in Norths
sociologically and empirically grounded theorizing.
But there was a divergence, too, which I was puzzling about until I later
heard Thomas Homer-Dixon speak at MIT in a seminar on environmental
stress, migration and violence. Homer-Dixon has a chapter in Global Accord,
a major research project underway, and considerable notoriety from Robert
D. Kaplans Malthusian look at Third World development prospects.(Kaplan,
1994) He passed out two papers, one summarizing his projects research
fndings, and the second discussing the ingenuity necessary to avoid the
violence-provoking ecological disasters in the making reported on in that
summary
29
. He is not the fatalistic pessimist Kaplan makes him out to be; on
the other hand, he insists that real changes (adaptations) have to be made
if environmentally focused degenerations are not to occur with increased
frequencies.
The frst divergence troubling me was one of modeling styles. Unlike Epstein
or Holland, Choucri has used either econometric modeling and nonlinear,
feedback laden, Systems Dynamics formulations. Territorial concepts fgure
in Choucri-North theories, and data-based models for separate countries are
the norm, but territory is not explicitly represented in her models. Structural
adaptation and environmental degradation are nicely handled by her non-
linear feedback loops (whose dynamics are so easily graphed by Systems
Dynamics/STELLA software), but there are no message-passing autonomous
agents in her models; she and Rothenberg and most other modelers in their
style are using econometrics or Systems Dynamics, not object-oriented
programming.
The missing link was migration! If Choucri and North were to model their
processes of lateral expansion on territorial grids, and these territorial grids
were to be like Epsteins sugarscapes, they would capture a new set of emer-
gent interaction possibilities. Ironically, some of Homer-Dixons best slides
were of highly polluting sugar or prawn plantations located on low level
fatlands at the foot of, or in between, Philippine mountains where much
poorer farmers held on to their degraded, non-terraced, much less attractive
production possibilities, or from which these farmers migrated to highly pol-
luted city slums in increasing numbers. Obviously a later version of Epsteins
software should have resource rich sugar spots on fatlands near the harbors,
not way up in the mountains. Various groups could migrate, trade and talk
among themselves and with others, not just view their unequally limited
horizons, etc. Message passing could also be used as a control process, to
calculate systems wide environmental limits processes as well (simulating
global warming, nonrenewable resource elimination, etc.) The resource rich
high points of his societies/civilizations would historically have been river
basins. A real shift to a more socially complex, historically adaptive multi-
agent simulation modeling style might well result.
Modeling.Historicity.on.the.Sugarscapes.of.Life
A growing impression throughout the meeting was that much of what I have
been working on as a modeler over the last 20 years could conceptually and
technically be ftted into what I called volume 2 of Epstein-Axtells project.
Object-oriented LISP routines could be imbedded in Epsteins object-oriented
PASCAL programs to explore new domains. This was a pregnant conver-
gence of projects par excellence. It would also help improve upon Choucri
style modeling, and transform it into the artifcially intelligent SCHEME/LISP
style of modeling that allows historicity-rich process representations
30
.
Let me further try to set up this line of thought by frst emphasizing a diver-
gence. The mechanisms featured in Hollands genetic algorithms (reviewed
in Mitchell, 1996) and Epsteins models are low down on the Lovejoy/
Boulding/Campbell Great Chain of Being/Acting/Knowing mentioned
above. Specifcally, collectively, socially, humans have evolved memory-rich,
language-rich, imagination-rich, organization-rich modes of problem-solv-
Alker
ingfrms, universities, markets, nations, and even science itselfwhose
mechanisms/processes/ capacities/ tendencies are not well represented in
either Hollands or Epstein-Axtells models. To put it bluntly, there is some
sociology, no cultural anthropology, no politics, no law, and no science in
Epsteins models. There are blue-dot-consuming fghts on his sugarscapes,
but not the games, and certainly none of the debates that Rapoport wrote
so well about (Rapoport, 1960).
As previously explicated, I think of the discovery of historicity as the birth
of the time ordered, [collective] self-understanding of a continuing human
society. Nations socialize their children into the exploits of their founding
heroes and great transitional actors, and the evils of their age old enemies,
with powerful effects: Le Pen is citing J oan of Arc to argue for throwing out
the foreigners, the Bosnian Serbs are still throwing out the Turks! A lot of
projecting of domestic enemies on to international antagonists relies on
this shared kind of mythical understanding. And since Auschwitz and Hi-
roshima, many are trying to create a new, anti-nuclear-war and anti-genocidal
historicity for world society.
Now the operational implication for Epstein, volume 2, of such kinds of
thinking would be to allow each of his actors individually and collectively,
to have memories of past interactions, so that the above kinds of historicity
could be represented and their developments studied. There would be both
forgetting, and sharing of such memories. The middle classes higher up
on the sugarscapes could develop shared understandings of the kinds of
interactions one typically has with those lower down the sugarscapes, or across
the valleys. Color coding of actors in such memories could have emergent
effects like Holland showed us. Even without mimicking complex natural
language capabilities of the Chomskyan sort, ethnic groups, classes, and ter-
ritorially structured states could be modeled, and their memories imperfectly
shared. Cultures, stereotypes, histories would be part of the sugarscape world.
Principles of political and territorial organization could be suggested, and
the identities, loyalties, and understandings typically associated with them
could be crudely, but suggestively modeled. A higher level of organizational/
jurisdictional/responsibility boundaries could be identifed (in a LISP-like
way on territorial grids) allowing for the tragedies of the commons style
ecology/security dynamics mentioned in the Santa Fe discussions. Historical
bases for imagining new collective responses to these challenges could also
be represented, and searched for, if memories were allowed different kinds
of empathetic, recreative ingenuity.
Obviously, the past paragraph requires much elaboration and much new
research before it is flled out. But I would like to signal here a strong con-
vergence between the programming styles of contemporary MIT artifcial
intelligence programming and that of Epstein, Holland, and other Santa
Fe researchers. Together they suggest what a new generation of intelligent
complex adaptive systems might look like. The key metaphor is the Piaget-
Pappert-Minsky one: the society of mind works by having differently capable,
nearly autonomous agents, which have their own computational capacities and
also pass messages to each other. Hierarchical aspects of thinking abound
as well, but in fexible and sometimes ad hoc ways. Memory (k-lines) is
also important, but limited. Minskys (1996) book presciently and correctly
conveys the programming style of some of his most illustrious students
(Winston, Hewitt, Sussman, Winograd, Stallman, Haase, Mallery, etc.) The
extension of this complex, merged style of modeling to the social domain is
very incomplete, but nonetheless promising.
31

Effective.Scenario.Making.Need.not.be.Model.Based
A third convergence at Santa Fe was, I think, not fully anticipated. When
I publicly repeated a private suggestion that our future thinking would be
scenario based, not model generated, a problem arose. What was the point of
modeling, anyway? One of our members correctly anticipated that scenario-
based future explorations would themselves generate model-based exercises,
not vice versa. With his own considerable experience in advising decision
makers, Dr. Zhenghua Jiangs accounts of the ways he presented different
scenarios in between best and worst possible cases generally supported the
view that modeling and future thinking could be symbiotically related. But
again we would not be able to derive the future from our models, or all rel-
evant alternatives.
32

The convergence of the practice of the workshop to such anticipations was
nearly exact. Thus Brian Arthurs worries about North-South style class
conficts, or subnational/transnational dual economies focused our think-
ing nicely. Nathan Keyftz has written about a global middle classitself a
model sketch developed in demographic, yet non-Malthusian, reaction to the
limits to growth debate (Keyftz, 1977). Murray Gell-Mann added brave new
worlds style class/race politics to this scenario. We could add the increased
or decreased legitimization of genocide and nuclear war. Nonetheless, the
Alker
heuristic role of the sugarscapes models could clearly be seen. Alternatives
to the Keyftz-Arthur-Gell-Mann scenario developed further in the conver-
sation; ideas for specialized, model-based research were generated. Even a
range of possible actor classes in the scenarios explored in CLAWs fnal
report were identifed.
None of the previous convergences can be taken for granted. None of the
ideas previously sketched is fully feshed out. Each is somewhat dependent
on the others. Nonetheless, I think these three convergences are suffciently
exciting to suggest that they imply new areas for future work, new areas of
collaboration in the social sciences domain, new ways of moving intelligent
complex adaptive systems research closer to the global realities of war and
peace. This is a line of thinking which Professor Seki has already begun
explore. I hope others will continue to do so, because I believe historicity-
rich modeling traditions better allow global citizens to learn from, and teach
about, the lessons of Hiroshima, Auschwitz, and Chernobyl as well.
Toward.Better,.Still.Revisable,.Ontologies
To help fulfll the emancipatory promise of research on achieving peace, the
operational concepts embodied in our intelligent complex adaptive systems
models need to be more intelligently constructed, working from richer ontolo-
gies of historical representational possibilities
33
. Simulations with affectively
encoded, symbolically represented memories need to contain ways of rec-
ognizing the practices constituting such past realities while at the same time
containing the possibility of transcending those realities through revisions in
their reality-codifying practices. As Hiroharu Seki has argued concerning the
signifcance of the end of the Cold War, we should include our political will
[,..our] moral leadership as one of the most signifcant variables....[W]ithout
the preference of ontology to epistemology or without the preference of sub-
jectivity to objectivity...we can not start the road towards solving any kind
of the global problematique.(Seki, 1996) Our ontologies of the real, the
possible, the desirable and the awful, the necessary and the impossible must
be revised to make such progressive development possible. I see Professor
Seki as someone trying to redress the balance in a world of social scientists
too committed to less adequate naturalistic ontologies.
Up until this point, except concerning the meaning of historicity, I have
used, but avoided detailed discussions of, ontological issues of primarily a
philosophical or metaphysical sort.
Unfortunately, many social scientists are embroiled rather unproftably in an
epistemological/ ontological debate between materialistic realists, ideal-
ists, and constructivists regarding essences and phenomenal appearances,
ontologies and epistemologies, natural and social realities. Some participants
in such debates might attempt, with some success, philosophically to unravel
and criticize the language of ontologies with which I have structured the
present discussion. But my focus here has been more operational, pointed
toward the design of historicity-friendly, war and peace relevant event rep-
resentations and intelligent complex adaptive systems. Therefore, let me
suggest briefy some exciting new directions in social scientifc research
grounded in more sophisticated conceptions of social reality than much of
the existing literature of International Relations, ontological conceptions
with which I feel positively identifed. These were completed or published
since the frst version of this text was written, yet they speak directly to the
main points I have made.
For example, Gavan Duffy has proposed reconstructing event data analysis
using the philosophical insights from W. V. O. Quine, Nelson Goodman,
and Hilary Putnam. He accepts Putnams notion of internal realism as a
preferred alternative to either a copy or a correspondence theory of truth,
and rejects both a copy theory or a constructivist theory of world-making.
Metaphorically, both the mind and the world jointly make up themselves.
Each person cognitively encodes a version of the world, which no one can
identify as objective truth, in the copy theory sense. For internal realists,
we each believe our versions right only in the sense and to the extent that
they cohere with our other ideas. These include our conceptual constructs,
our internalized norms, our perceptual felds, our life-experiences, and even
our mental models of the world versions that those with whom we sociate
disclose to us in discursive and practical action (Duffy, 1996, p. 148f). Duffy
then goes on to suggest principals for a representationally sophisticated, yet
scientifcally replicable, computational hermeneutics, based signifcantly on
his years working with J ohn Mallery and others in the development of the
RELATUS system.
In a remarkably similar vein, the philosophical computer scientist Brian
Cantwell Smith has proposed a middle ground between materialist and ide-
alist metaphysics, a ground conceived of at least partly in terms of revisable
0 Alker
computational ontologies, the sustained metaphor of this present discussion.
He suggests a philosophy of presence governed by:
symmetrical realism: a commitment to One world, a world with no other,
a world in which both subjects and objectsare accorded appropriate
place[T]hatworldisdepictedasoneofcosmicandultimatelyineffable
particularity:acriticallyrichandall-envelopingdeictic[situationreferring]
fux[which]sustainscomplexprocessesofregistration:aformofinterac-
tion,subsumingbothrepresentationandontology[But]registration[has]
metaphysical zest: the fact that all distinctions and stabilitiesempirical,
conceptual, categorical, metaphysical, logicalare taken not, at least not
necessarily,andnotinthefrstinstance,tobeclearanddistinct,sharp,
or in any other way formal, but instead to be, or at least potentially to be,
wily, critical, obstreperous, contentious, and in general richly eruptive with
fnestructure(Smith,1996,p.347f).
Here is a world uniting phenomenologies and ontologies, with plenty of room
for moral imperatives, political and ontological differences, and inter-sub-
jectively shareable historical learning, a world that Smith is suggesting we
replicably explore with computationally facilitated symbolic procedures.
Finally, I call to my readers attention to an impressive (yet horror flled),
phenomenologically sophisticated, LISP-encoded account of many the
overlapping and intersecting political worlds of what Russians usually call
The Caribbean Crisis and Americans call The Cuban Missile Crisis.
Schmalberger suggests that this crisis was about humiliation, prestige
and leadership. It was a drama played before the world in which deception
commonly used to enhance ones own prestige at the expense of the other,
threatened to destroy the face of the protagonists.(Schmalberger, 1998, p. 5)
Without reducing these interactions to causal models, Schmalberger provides
text-based, phenomenologically grounded reconstructions of 66 conversa-
tion-like turns in the crisis, and analyzes this sequence counterfactually,
generating alternative possible scenarios which lead the historically minded
to many sobering refections about the rationalistic self-misunderstandings
of the key players.
These counterfactual reconstructions of the political process help facilitate
deep learning about what brought the world the closest it has come to nuclear
war since Hiroshima and Nagasaki. They suggest ways of behaving that
should become part of the historicities of international and world society.
Social and scientifc international learning regarding nuclear war avoidance
requires such historicities. If the next generation of International Relations
scholars and intelligent complex adaptive systems modelers can build further
on these fne ontological contributions to an ecologically sustainable, nuclear-
war avoiding future, Professor Seki and I should be content.
References
Alker, H. R. (1997). Putting historicity into adaptive, multi-agent simula-
tions: So that Hiroshima and Auschwitz might be remembered but not
repeated. Ritsumeikan Journal of International Affairs, 9(4), 125-132.
Alker, H. R. (1996, digitally reprinted 2007). Rediscoveries and reformula-
tions: Humanistic methodologies for international studies. New York:
Cambridge.
Alker, H. R., & Frasier, S. (1996). On modeling historical complexity:
Naturalistic modeling approaches from the Santa Fe Institute. Paper
presented on August 31 at the American Political Science Association
meeting.
Alker, H. R., Gurr, T. R., & Rupesinghe, K. (2001). Journeysthroughconfict:
Narratives and lessons. Lanham, MD: Rowman & Littlefeld.
Alker, J r., H. R. (1966). The long road to international theory: Problems of
statistical nonadditivity. World Politics, 18(4), 623-655.
Alker, Jr., H. R. (1988). Bit fows, rewrites and social talk: Towards more
adequate informational ontologies. TodaiSymposium86,Information
and its Functions, Tokyo: Institute of J ournalism and Communications
Studies, University of Tokyo.
lker, J r., H. R. (1987). Against historical/disciplinary genocide as a prereq-
uisitetohistorysscientifcdissection:Puttinghistoricityatthecenter
of events data research. Presented at the 1987 meeting of the American
Political Science Association.
Alker , Jr., H. R., & Brunner, R. D. (1969). Simulating international confict:
A comparison of three approaches. International Studies Quarterly,
13(1), 70- 110.
Alker
Alker, J r., H. R., Duffy, G., Hurwitz, R., & Mallery, J . C. (1991). Text model-
ing for international politics: A tourists guide to RELATUS. (1991). In
V. Hudson (Ed.), Artifcialintelligenceandinternationalpolitics (pp.
97-126). Boulder, CO: Westview.
Axelrod, R. (1997). The complexity of cooperation: Agent-based models
of competition and collaboration. Princeton, NJ : Princeton University
Press.
Bennett, J . P., & Alker, J r., H. R. (1977). When national security policies bred
collective insecurity: The War of the Pacifc in a world politics simula-
tion. In Deutsch, Fritsch, J aguaribe and Markovits (1977), 215-302.
Bremer, S. A., & Mihalka, M. (1977). Machiavelli in Machina: Or politics
among hexagons. In Deutsch, Fritsch, J aguaribe & Markovits (1977),
303-338.
Bull, H. (1977). The anarchical society. New York: Columbia University
Press.
Buzan, B., & Waever, O. (2003). Regions and powers: The structure of in-
ternational security. Cambridge: Cambridge University Press.
Cederman, L. E. (2001). Modeling the Democratic Peace as a Kantian Selec-
tion Process. JournalofConfictResolution,45(4), 470-502.
Cederman, L. E. (1997). Emergent actors in world politics: How states
and nations develop and dissolve. Princeton, NJ : Princeton University
Press.
Cederman, L. E., & Daase, C. (2003). Endogenizing corporate identities: The
next step in constructivist IR theory. European Journal of International
Relations, 9(1), 5-35.
Choucri, N. (1993). Global accord: Environmental challenges and interna-
tional response. Cambridge, MA: M.I.T. Press.
Choucri, N., & North, R. (1975). Nationsinconfict. San Francisco: W. H.
Freeman.
Choucri, N., North, R., & Yamakage, S. (1992). The challenge of Japan:
Before World War II and after. London & New York; Routledge.
Cohen, P. R., Morgan, J ., & Pollack, M. E. (1990). Intentions in communica-
tion. Cambridge: M.I.T. Press.
Cohen, P. R., Schrag R., J ones, E., Pease, A., Lin A., Starr, B., Gunning, D.,
& Burke, M. (1998). The DARPA [U.S. Defense Advanced Research
Projects Agency] high-performance knowledge bases project. AI Maga-
zine, 19(4), 25-49.
Cusack, T. R., & Stoll, R. J . (1990). Exploring Realpolitik: Probing interna-
tional relations theory with computer simulation. Boulder, CO: Lynne
Rienner.
Cusack, T. R., & Stoll, R. J . (1994). Collective security and state survival in
the interstate system. International Studies Quarterly, 38(1), 33-60.
Dallmayr, F. (1991). Between Freiburg and Frankfurt: Toward a critical
ontology. Amherst: University of Massachusetts Press.
Der Derian, J . (1998). The Virilio reader. Oxford: Blackwell.
Der Derian, J . (1991). Anti-diplomacy: Speed, spies, and terror in interna-
tional relations. Basil Oxford: Blackwell.
Der Derian, J ., &. Shapiro, M. J . (1989). International/intertextual relations:
Postmodern readings of world politics. Lexington, MA: Lexington
Books.
Deutsch, K. W., Burrell, S. A., Kann, R. A., Lee, J r.,, M., Lichterman, M.,
Lindgren, R. E., Loewenheim, F. L., & Van Wagenen, R. W. (1957).
Political community in the North Atlantic area. Princeton, NJ : Princeton
University Press.
Deutsch, K. W., Fritsch, B., J aguaribe, H. & Markovits, A. S. (1977). Prob-
lems of world modeling: Political and social implications. Boston:
Ballinger.
Duffy, G. (1994). Events and versions: Reconstructing event data analysis.
In (Duffy, ed, 1996), 147-167.
Duffy, G., Guest Editor, (1994). New directions in event data analysis. In-
ternational Interactions, 20(1- 2), 1-168.
Duffy, G. (1992). Concurrent interstate confict simulations: testing the ef-
fects of the serial assumption (in Cusack-Stolls model). Mathematical
and Computer Modelling, 16(8/9), 241-270.
Duffy, G., Frederking, B. K., & Tucker, S. A. (1998). Language games: Dia-
logical analysis of INF negotiations. International Studies Quarterly,
42(2), 271-294.
Dunn, T. (1998). Inventing international society: A history of the English
school. Houndmills and New York: Macmillan and St. Martins.
Alker
Epstein, J., & Axtell, R. (1996). Growing artifcial societies: Social science
from the bottom up. Washington, DC: Brookings.
Gell-Mann, M. (1994). The Quark and the J aguar: Adventures in the simple
and the complex. New York: W.H. Freeman.
George, J . (1996). Understanding international relations after the Cold War.
In (Shapiro, M. J . & Alker, H. R., eds. (1996, 33-79).
Gerner, D. J ., Schrodt, P. A., Francisco, R. A., & Weddle, J . L. (1994). Ma-
chine coding of event data using regional and international sources.
International Studies Quarterly, 38(1), 91-120.
Gordon, A. (2004). Strategy representation: An analysis of planning knowl-
edge. Mahwah, NJ : Lawrence Erlbaum.
Hammond, A. (1998). Which world? Global destinies, regional choices.
Washington, DC: Island Press/Shearwater Books.
Holland, J . H. (1995). Hidden order: How adaptation builds complexity.
Reading, MA: Addison Wesley.
Homer-Dixon, T. (2000). The ingenuity gap. New York: Knopf.
Homer-Dixon, T. (1995). The ingenuity gap: Can poor countries adapt to
resource scarcity? Population and Development Review, 21(3), 587-
612.
Homer-Dixon, T. (1994). Environmental scarcities and violent confict: Evi-
dence from cases. International Security, 19(1), 5-40.
Hopf, T. (1998). The promise of constructivism in international relations
theory. International Security, 23(1), 171-200.
Kaplan, R. D. (1994). The coming anarchy. Atlantic Monthly. February 27
issue, 44-75.
Keyftz, N. (1977). World resources and the world middle class. In (Deutsch,
Fritsch, J aguaribe & Markovitz, 1977), 97-116.
King, G., & Zeng, L. (2001). Improving forecasts of state failure. World
Politics, 53(4), 623-658.
Lakoff, G., & Johnson, M. (1999). Philosophy in the fesh: The embodied
mind and its challenge to western thought. New York: Basic Books.
Lasswell, H. (1950, originally 1934). World politics and personal insecurity.
Reprinted in H. D. Lasswell, C. E. Merriam, & T. V. Smith (Eds.), A
study of power. Glencoe, IL: Free Press.
Lustick, I. S. Miodownik, D., & Eidelson, R. J . (2004). Secessionism in mul-
ticultural states: Does sharing power present or encourage it? American
Political Science Review, 98(2), 209-230.
Lyotard, J . F. (1984). The postmodern condition: A report on knowledge.
Minneapolis, MN: University of Minnesota Press.
Merritt, R. L., Muncaster, R. G., & Zinnes, D. A. (1993). International
event-data developments: DDIR Phase II. Ann Arbor, MI: University
of Michigan Press.
Minsky, M. (1987). Society of mind. New York: Simon and Schuster.
Mitchell, M. (1996). An introduction to genetic algorithms. Cambridge, MA:
M.I.T. Press.
Olafson, F. (1977). The dialectic of action: A philosophical interpretation of
history and the humanities. Chicago: University of Chicago Press.
Rapoport, A. (1960). Fights, games, debates. Ann Arbor, MI: University of
Michigan Press.
Roig, C. (1980). La grammaire politique de Lnine, Lausanne: LAge
dHomme.
Schelling, T. (2006). Micromotives and macrobehavior. New York: Nor-
ton.
Schmalberger, T. (1998). Dangerous liaisons: A theory of threat relation-
ships in international politics, unpublished Ph.D. Thesis #581, Institut
Universitaire de Hautes tudes Internationales, Genve.
Seki, H. (1996). How nuclear weapons were related to the end of the Cold
War. Paper presented on September 17 to the Hiroshima Symposium
on The Fifty Years of Nuclear Weapons, Hiroshima.
Shapiro, M. J ., & Alker, H. R. (1996). Challenging boundaries: Global
fows,territorialidentities. Minneapolis, MN: University of Minnesota
Press.
Smith, B. C. (1996). On the origin of objects. Cambridge, MA: M.I.T.
Press.
Sylvan, D. A., Ostrom, T. M., & Gannon, K. (1994). Case-based, model-
based, and explanation-based styles of reasoning in foreign policy.
International Studies Quarterly, 38(1), 61-90.
Sylvan, D., & Thorson, S. J . (1992). Ontologies, problem representation,
and the Cuban Missile Crisis. Journal of Confict Resolution, 36(4),
709-732.
Alker
Wight, C. (2006). Agents, structures and international relations: Politics as
ontology. Cambridge: Cambridge University Press
Endnotes
1
I wish to dedicate this essay to the memory of Professor Hiroharu Seki,
late of Ritsumeikan University. Without holding them responsible for
their contents, I would like to acknowledge the contributions to these
refections of a sabbatical visit to the Santa Fe Institute (SFI), and related
conversations with W. Brian Arthur, Robert Axelrod, Nazli Choucri,
J oshua Epstein, Murray Gell-Mann, J ohn Holland, Stuart Kauffman,
Christopher Langton, George Lakoff, Melanie Mitchell, and Thomas
Ray. Not at the SFI, Thomas Schmalberger has also been a great inspira-
tion. An earlier version of this essay was SFI Working Paper 99-02-011,
which in turn was an expanded rewrite of (Alker, 1997)
2
One of the broader and more thoughtful treatments of constructivism
in the war/peace context is Hopf (1998). For a British variant of social
constructivism, see Tim Dunnes excellent historical-analytical account
(Dunn, 1998). For a continentally oriented discussion of ontological
themes reminiscent of some of the issues raised during SFI workshops
on economics and cognition see Dallmayr (1991). An excellent recent
modelers treatment of social constructivism, relying heavily on Georg
Simmel, is Cederman and Daase (2003).The recent ontological turn in
IR theory is best articulated in White (2006).
3
A relevant, simplifed rehearsal of the classical plot structures in
contemporary historical accounts is (Knutsen, 2002); on post-modern
ironic sensibilities see especially (Lyotard, 1984). In a discussion of the
corrupted, nightmarish consequences of the Enlightenments narrative
of reason, knowledge and freedom, J im George suggests we connect
the ascent of the modern, rational subject with the experiences of Hi-
roshima and Auschwitz. (George, 1996, p. 54) Without denying the
connectionsJohn von Neumann fts into both narrativesI should state
my view that the Enlightenments scientifc and political contributions,
although fawed, are far too important and valuable to be summarily
dismissed.
4
For example, I cite from an exemplary study by one of the best contem-
porary complex adaptive systems modelers in International Relations
today: [A] variable-sized grid constitutes the core of the model
[T]he main simulation loop contains fve phasesAs in my previous
models, the logic of execution is quasi-parallel with double buffering
To achieve quasi-parallel execution, the list of actors is scrambled each
time structural change occurs. The actors keep a memory of one step
and thus make up a very complex Markov-like process (Cederman,
2001, p. 493). A similar analysis could be made of the frst complex
adaptive systems model featured (a multi-colored rectangular checker-
board-like grid on the front cover, no less) in the American Political
Science Review: (Lustick, Miodownik, & Eidelson, 2004).
5
Event data are typically nominal, ordinal or interval scale measurements
of the kind or degree of cooperation or confict between two parties,
based on media accounts. For a scientifcally respectable review and
extension of this literature see Merritt, Muncaster, and Zinnes (1993)
6
The frst and most intellectually taxing task when building a large
knowledge base is to design an ontology. From Cohen et al. (1998, p.
26). Sylvan and Thorson (1992) agree on the importance on ontologi-
cal issues, arguing that different ontologies can be more important than
options analysis in determining decision-making possibilities in the
Cuban Missile Crisis.
7
Adaptive, multi-agent modeling in International relations is usually said
to start with Bremer and Mihalka (1977). But this book also contains
a more intelligent, empirically corroborated, memory-rich, geographi-
cally structured, adaptive, multi-agent predecessor of Cedermans and
Duffys more recent work: Bennett & Alker (1977). This article, based
on Bennetts prize-winning thesis, might be said to pioneer intelligent
complex adaptive systems in International Relations. Robert Axelrod
and his student Lars-Erik Cederman are also rightly seen as pioneers in
this style of International Relations modeling. See Axelrod (1997) and
Cederman (1997). The most important recent theoretical refnement of
this literature, paralleling the Bennett-Alker chapter in several respects,
is: Lars-Cederman and Daase (2003).
8
Alker, J r., H. R. & Brunner, R. D. (1969). More elaborated descrip-
tions of, and bibliographic citations to, the TEMPER, INS and Political
Military Exercise simulations are given in that article.
Alker
9
The post-structuralist/post-modern discussion of ambiguous, mis-
leading, simulated international events by Baudrillard and Virilio was
introduced into academic International Relations in Der Derian and
Shapiro (1989) and Der Derian (1991, 1998).
10
See the contributions to Merritt, Muncaster, and Zinnes, op. cit., and
the historically enriched proposals in Alker, Gurr, and Rupesinghe
(2001).
11
This section presents, in abbreviated and revised form, the substance of
my provocatively entitled memo (Alker, 1987). Because no direct acts
of violence are committed by such defcient coding practices, and the
premeditations involved are not usually murderous, I would now retitle
that memo and here propose for such a retitling the terms: Cliocide
(and a related adjective Cliocidal), etymological references to the
killing of the mythical Goddess of History, Clio. See the text below for
a fuller defnition of the silencing or symbolic killing of collective
historical-political or disciplinary identities and identifying practices
by historical- or discipline- defcient scientifc coding practices.
12
See chapters 5 (Can the end of Power Politics...) and 10 (Emancipa-
tory Empiricism) of (Alker, 1996).
13
See Alker (1996, passim), Schmalberger (1998), and Alker, Gurr, and
Rupesinghe (2001).
14
The relevant citation for this conception of historicity is Olafson (1977),
which also richly reviews alternative conceptions of what sometimes is
called the essence of history. Further discussion of this seminal, partly
Heidiggerian, partly Marcusian reformulation is contained in Chapter
11 of Alker (1996).
15
The importance of regional contexts of meaningsand the relevance of
covariance and regional factor analyses for multiplicatively reformulating
relevant operational hypotheses and conceptual indicatorsis an often
neglected component of contemporary statistical training in International
Relations. See Alker (1966). The study of regional security complexes
has recently been improved upon in Buzan and Waever (2003).
16
Sources for these interpretations can be found in Alker (1996, Chapter
1).
17
A clarifcation of this ontological perspective and some earlier thoughts
on informational and historical ontologies are offered in Alker (1988).
See also the pioneering discussions in (Duffy, 1994). Reconstructing the
metaphorical (and hence not metaphysically necessary) bases of much
ontological thinking about human affairs is a principal achievement of
Lakoff and J ohnson (1999).
18
Although I cant fnd an exact source, this summarizing text is doubtless
inspired by Harold Lasswell (1950). See especially his Chapter 1: The
confgurative analysis of the world value pyramids; Chapter 2: Nations
and Classes: The symbols of identifcation; and Chapter 3: The Balanc-
ing of Power: The expectation of violence.
19
See Karl W. Deutsch et al. (1957), Bull (1977), and Roig (1980).
20
Besides the various articles in Duffy (1994), there are several practically
oriented efforts within Political Science and International Relations to
get increased purchase on the voluminous body of daily newspaper ac-
counts. For a thoughtful but modest effort to use linguistically-informed
methods of event characterization, see Gerner, Schrodt, Francisco, and
Weddle (1994). More ambitiously, see the description and analysis of
the U.S. Central Intelligence Agencys State Failures project published
by its statistical critics: (King & Zeng, 2001).
21
An excellent, Artifcial Intelligence-inspired paper on such conversa-
tional modalities of practical reasoning is Sylvan, Ostrom, and Gannon
(1994).
22
The chapters by Mallery and Duffy in Duffy, ed., op. cit., are the most
relevant discussions of related procedures. See also Alker, Duffy, Hur-
witz, and Mallery (1991); and the more dialogically oriented Duffy,
Frederking, and Tuckner (1998).
23
The rest of this section recasts a memo prepared originally in April 1994
for the Crude Look at the Whole project at the Santa Fe Institute,
and reported on previously in Alker 1997. Its original title was Some
Pregnant Convergences Observed at the CLAW Workshop. Both its
composition and my visit to SFI were centered around the working out
of my concern with war-avoiding historicities, the most central redis-
covery of my earlier research, as represented and further developed
in Alker (1996) and Alker and Fraser (1996). The most relevant, acces-
sible books of these speakers are Holland (1995) and Epstein and Axtell
(1996).
24
If an ECHO model is completed through the use of the SWARM
modeling platform developed at SFI by Christopher Langton and his
associates, or the REPAST modeling platform Cederman now prefers,
0 Alker
this may argue for trying to introduce richer historical capacities into
ECHO rather than the Epstein-Axtell Sugarscape model. See Alker
and Fraser (1996) for more details.
25
Besides the Bremer and Mihalka work cited in End Note 8, see Cusack
and Stoll (1990, 1994) and Duffy (1994). Of course, also see the earlier
versions of Schelling (2006).
26
A good over all introduction to Gell-Manns vision of SFI research is
Gell-Mann (1994).
27
See the discussion and references in Alker (1996, p. 98-100).
28
Other relevant citations to Choucri and her collaborators are included
in this volumes Bibliography.
29
The book length development of this later argument is Homer-Dixon
(2000).
30
For example, see Schmalberger (1998). An overview and demonstrative
run of this simulation is available on http://www.usc.edu/dept/LAS/ir/
cis/cews/html_pages/Dangerous_Liaisons/index.html.
31
At some levels of linguistic, social, and historical complexity, such
mergers become too complex to handle. Then, moving to the intelligent,
adaptive modeling of symbolic cognition becomes a possible alternative.
In this regard Cohen, Morgan, and Pollack (1990) and Gordon (2004)
come readily to mind. But also review Duffy and Mallerys M.I.T.
work on RELATUS, Bennetts War of the Pacifc simulation, and the
Cusack-Stoll and Duffy remodelings of Bremer and Mihalkas early
Machiavelli in Machina.
32
The related SFI sponsored volume is Hammond (1998).
33
It is interesting in this regard that Cedermans excellent complex adap-
tive systems model of the coming of the democratic peace identifed in
End Note 5 does not explicitly model the more complex processes of
community building and community disintegration, focusing rather on
forceful conquests and calculated alliance behaviors.
The Allocaton of Complexty n Economc Systems
Chapter.XI
The.Allocation.of.
Complexity.in.
Economic.Systems
1
J ason Potts, Unversty of Queensland & Queensland Unversty of
Technology, Australa
Kate Morrison, Volterra Pacifc Pty. Ltd., Australia
J oseph Clark, Unversty of Queensland & Suncorp Pty. Ltd.,
Australa
Abstract
This chapter isolates a classic allocation problem in the substitution relation
between two primary carriers of complex rulesagents and institutionsas
a function of the relative costs of embedding rules in these carriers, all sub-
ject to the constraint of maintaining overall system complexity. We call this
generic model the allocation of complexity, which we propose as a bridge
between neoclassical and complexity economics.
Potts, Morrson, & Clark
Introduction
The market economy is an excellent example of an intelligent complex adap-
tive system. Economists, however, have largely failed to develop theoretical
frameworks based upon this insight and have instead persisted with a time-
less equilibrium-based analysis. The exception to this has been the work of
the Austrian economists (such as Friedrich Hayek) and evolutionary econo-
mists (such as J oseph Schumpeter). In this chapter, we propose a novel way
of analyzing the complexity of self-organization in economic systems that
draws upon the key model of the equilibrium-based neoclassical framework
in terms of a comparative static framework for analysis of the allocation of
rules between different classes of carrier. We start by assuming complexity
(i.e., that evolutionary forces maintain complexity in open systems) and
then analyze the distribution of state-space equilibria under different relative
costs/prices of embedding rule-complexity in different carrier systems, such
as agents or institutions.
The outcome is an allocation of complexity. Changes in relative prices, as
caused by technological, institutional, or fnancial innovation, say, will ef-
fect the position of the equilibria in carrier-space. We may, therefore, study
how change in the cost of embedding rules conditions the evolution of the
complexity of an economic system. The upshot is a framework for arraying
Figure 1. The allocation of complexity model of rules in agents and institutional
carriers as a function of the relative price of embedding isocomplexity

Complexity of carrier X
Complexity
of carrier Y
Isocomplexity curve
Relative price of embedding
complexity in |Y/X|
Equilibrium
allocation of
complexity
economic forces over the different dimensions of economic evolution. This
enables us to study the effects of relative price changes on the dynamics of
economic evolution. The use of complexity as a conservational concept (and
its derivative, a convexity argument) may, we suggest, prove expedient for
further theoretical and conceptual development of economic analysis. The
generic graphical model, which we propose as a possible bridge between
neoclassical and complexity economics is as such:
In modern evolutionary economics (Dopfer, Foster, & Potts, 2004a, Dopfer
& Potts, 2007), an economic system is conceptualized as a system of meso
units, with each meso unit the population of actualizations (or carriers) of a
generic rule. Economic evolution is then defned as a process of change in
the meso population of rules. The micro-meso-macro analytical framework
reveals the underlying structure of economic evolution as a meso trajectory,
and frames analysis of this process in terms of micro and macro domains.
However, no account has yet been given of the economic nature of the
complexity in this process, and, in particular, of the allocation problem of
rules over carriers. What, in other words, are the economics of economic
evolution? Evolutionary economists normally emphasize complementarity
and connective structure (e.g., Potts, 2000). But here we propose that eco-
nomic evolution can also be considered in terms of substitution and relative
prices. This is achieved by recasting the prime dimensioning of analysis in
terms of the comparative rule-complexity of agents and institutions, or, more
generally, of carriers.
The economic system is a rule-system and the economic process is a rule-
process. From this premise, we may then isolate a classic allocation problem
in the substitution relation between two primary carriers of complex rules
agents and institutionsas a function of the relative costs of embedding rules
in these carriers, all subject to the constraint of maintaining overall system
complexity. The allocation of complexity problem is, therefore, defned as
analysis of changes in the relative costs of rule embedding in different carriers
(agent and the environment) on the (allocative) distribution of complexity
between them. Our fundamental premise is that complexity is carried in two
fundamental classes of carrieragents and institutionsand that difference
in the cost of embedding rules in each determines the equilibrium allocation
of complexity.
A framework of comparative statics, familiar to all students of economics,
may thus be used to analyze the distribution of complexity carried by an
economic system and how this changes as an evolutionary process. Section
2 introduces the relationship between rules and complex systems; Section 3
discusses the idea of complexity as a solution concept; Section 4 represents
this in terms of the allocation of complexity problem; Section 5 completes
the framework with the construct the relative price of complexity; Section
6 concludes.
Rules.and.Complex.Systems
The economic system is a complex rule-system. The conceptualization of
generic rule evolution is set out in Dopfer and Potts (2004b), and the micro-
meso-macro framework is presented in Dopfer et al. (2004a) and Dopfer et
al. (2007). A rule is a generic idea: a meso unit is then defned as a generic
rule and the population of.carriers of the rule. Beneath the surface phenomena
of production and exchange, there lies the deep structure of knowledge as a
complex system of generic rule populations; this is the abstract reality of the
economic system as a complex structure of knowledge. Our question then
arises: what determines the structure of this knowledge base of differentially
embedded rules?
Economic evolution is the creative-destructive process of the origination
and embedding of new rules and the displacement of others. This is analyti-
cally defned as a process of (generic) meso coordination and change along a
meso trajectory of origination, adoption, and retention that occurs as a new
technology (i.e., a generic rule) enters and changes the knowledge base of the
economic order. The meso-centered analytical framework allows us to model
economic evolution as the self-transformation of a complex rule-system. It
links together evolutionary economic analysis of population dynamics and
structural changeor agent learning and emergence of institutionsinto a
unifed framework. By connecting population and statistical arguments to-
gether with arguments about complex structure, it provides a foundation for
analysis of evolutionary processes as based on the coordination of complex
systems. It is both ontologically and analytically coherent. Yet the generic
meso-based model leaves the economics of this process unnecessarily, we
think, unstated. Specifcally, a basic assumption of the micro-meso-macro
framework of economic evolution is that rules are embedded in the eco-
nomic system in carriers. Yet Dopfer et al. make no attempt to tease out the
specifc form in which they are embedded, or to account for the operational
economic forces that effect the allocation of rule complexity over different
classes of carriers.
We propose, then, that the allocation of complexity framework can be used
to analyze the rule substitution processes interior to a meso trajectory. A rule
can be carried or, equivalently, embedded, in different ways: which is to say
that the complexity of the economic system can manifest in various structural
forms.
3
A technology, for example, could be embedded largely in the cognitive
abilities of human agents, or largely in commodities or institutions. Smarter
agents enable the same level of complexity to be achieved in a simpler en-
vironment, and a smarter environment may achieve the same with simpler
agents. For example, the generic rules of bookkeeping and accounting could
be embodied in accountants heads, as it were, or in interactive software, or
in some convex combination of the two carrier forms. But what determines
the equilibrium allocation at a point in time, or the path an evolutionary tra-
jectory of technology adoption will take? What determines this distribution,
in other words, of rule complexity over classes of carrier?
4
The allocation of complexity problem concerns the distribution of the popula-
tion of a generic rule over different classes of carrier as a function of relative
cost of adoption and embedding of the rule. Rules carried exclusively in one
class of carrier are, obviously, not subject to the allocation of complexity
problem. But, if carried in this way, they are also unlikely to be generic eco-
nomic rules in the frst place. Following Dopfer (2004) and Foster (2006),
we plainly recognize that economic systems are human-centric (a method-
ologically individualist stance) in that they are composed of structures of
knowledge (generic rules) that may be carried in agents or embedded in the
institutional environment to which the agent is subject.
Economists, along with social scientists and philosophers, have traditionally
drawn a sharp duality between rules carried in the mind or as behavior (by
subjects) from rules carried in organizations of matter-energy (e.g., a physi-
cal technologies or commodities) or organizations of people (e.g., a frm or
household). In the framework proposed here, these alternative carrier classes
are conceptualized in a continuous substitution relation along the line of
constant complexity (the isocomplexity function).
The agent-environment dichotomy is also a fundamental construct in com-
plex systems analysis, including the defnition of open and closed systems
and of concepts such as feedback. But we suggest that an allocation prob-
lemthat is, an economic problememerges only when choice over where
rules are embedded also exists. This is not the case for biological systems
whose agentsgenes, organs, butterfies, etc.are constrained completely
by their environment, precisely because these agents cannot exercise generic
choice over the rule-composition of their environment. While it is plain that
perhaps most organisms build some rules into their environment (e.g., bees
nests, beaver dams, etc.), economic agents build complexity into their en-
vironment to an extent orders of magnitude greater (Raine, Foster, & Potts,
2006). Indeed, as Foster (2006) explains, such higher-order complexity is
an effective defnition of an economic system.
Although the science of economics relies fundamentally on the notion of
human choice, we posit that the proper analysis of economic systems (or
other human-based emergent complex systems, be they social, political or
cultural) must rest on the recognition that human agents have the unique
capacity to externalize rules by sometimes choosing to embed them in their
environment.
5
It is the systemic consequences of such rule-embedding choices
that we herein seek to illuminate. We suggest, in turn, that this hybrid meth-
odology offers a useful way of unpacking a key mechanism in economic
evolution. Our starting axiom, then, is that an economic system is a complex
natural-artifcial system of high-order complexity that is most immediately
manifest in the variety of structural forms that a rule can be carried. The al-
location of complexity problem thus arises as economic forces (and not just
technological considerations) effect the equilibrium distribution of generic
rules over classes of carriers.
Rules, in general, are fungible with respect to carriers, of which there are many
types and classes. For ease of exposition and analytic parsimony, however,
we shall presume just two classes of carrier: agent and environment. We
shall assume that the complexity of a generic rule can be decomposed over
these two carrier classes. Within this, and in order to allow dimensionality,
each carrier class can assume varying levels of complexity. An agent, for
instance, can be very simple or highly complex. The environment is also pre-
sumed to vary in complexity between states of being very simple and highly
complex. While these two dimensions are clearly defned on a qualitative
metric, we have in mind a quantitative metric based on an entropy measure
of complexity in terms of energy or information processing as operation-
ally computed/processed by the rule system (Raine et al., 2006). We thus
infer (indeed axiomitize) a computational ontology of economic knowledge
(Mirowski, 2007; Potts & Morrison, 2007).
The set of all combinations of a generic rule over carriers is mapped by what
we call carrier-space. The allocation of complexity is of a generic rule over
carrier space. At a point in time, the complexity of a generic rule is carried
in one part by the complexity of the agent in terms of its generic inheritance,
generic learning, and operational acquisition, and in the other part by the
complexity of the environment in terms of institutions, technology, and capi-
tal. Economic evolution is the process of change in the generic rules of the
economic system: but where do these changes operationally manifest, and
why? What, in other words, is the economic explanation for the allocation
of the structure of knowledge over different carrier forms?
This question has general signifcance to the study of all intelligent complex
adaptive systems, and we should not be at all surprised to fnd that economic
systems, considered abstractly, offer insight into the nature of such systems.
An economic system is a complex system of knowledge (or generic rules,
in the language of our theory) composed of both rational and imaginative
agents that live by and grow in wealth to the extent that they can off-load
knowledge into the environment. Indeed, the nature and causes of the wealth
of nations can reasonably be summarized as the extent to which we have been
able to embed ideas and knowledge into the environment as technologies,
capital, and institutions. Yet this is always an ongoing process from which
every new idea (generic rule) begins as an idea in an individuals mind. A
technology, indeed, is the name we give to such an idea that has completed
the trajectory of adoption and retention into many agent carriers (as a meso
trajectory). But this process has multiple equilibrium outcomes along a con-
tinuum of carrier confgurations. Our question, thus, is on the determinants
of this distribution.

Assume.Complexity
Along with a specifc understanding of the economic system as a rule-system,
our theory is also based on a specifc analytical assumption about the nature
of complexity as a solution concept. Complexity is a partial region of the
potential state-space of a system, and is therefore a property that a system
may or may not have. A system may occupy one of many positions in the
state-space continuum, yet only some of these states are complex. What is
interesting about these states is that complexity seems to be an attractor.
This point has been made by Kauffman (1993) and others (e.g., Potts, 2000,
2001) in relation to the balance of forces of variation and selection, and by
Prigogine (2005) and others (e.g., Allen, 1998; Foster, 2006; Schneider & Kay,
1994) in relation to the second law of thermodynamics. As the deep forces of
energy mechanics in closed systems work to maintain an equilibrium state,
the deep forces of open system thermodynamics and evolution seem to work
to maintain complexity. The founders of the theory of fully connected market
systems (e.g., Walras, Edgeworth et al.) saw ft to suppose the long-run state
of the competitive economic system is general equilibrium, which became,
in turn, the starting place for standard economic analysis. In contrast, we ar-
gue that the competitively maintained long-run state of a partially connected
open-market economic system is general complexity, which is assumed to
be a general solution concept for an open system.
6
This is why we assume complexity, and therefore allow a differential
distribution of complexity over classes of carriers of a generic rule. Both
Kauffmans hypothesis of evolution toward complexity and the Kay and
Schneider hypothesis of complex systems maximally degrading energy
both refer to deep forces that maintain a state of complexity in an open sys-
tem process. There is mounting evidence that real world economic systems
exhibit small world and scale-free properties,
7
which thus reinforces the role
of complexity as a structural attractor in the state-space of a system. There
are, of course, many reasons that a system will not be complex, such as to
lock-in important structural features in the short run. But when this happens,
fexibility or adaptive potential is lost somewhere, which tends to eventu-
ally impact on the long run prospects of the generic rule. In an open system,
where there is always change in the fows of energy and information from
the environment, and where the environment itself is subject to change, we
should expect that the basic meso structure of an economic system will be
structurally complex. Economic systems are complex because, in essence,
they could not have evolved if composed otherwise (see Simon, 1969). The
forces that determine the allocation of this complexity should be of signif-
cant concern not just to economists, but to all students of intelligent complex
adaptive systems.
The.Allocation.of.Complexity
In consequence, the embedding dimensions of generic rules (i.e. technology
or knowledge) is a much neglected dimension of economic analysis. The al-
location of complexity framework seeks then to fll this gap by proposing a
decomposition of the distribution of rule complexity between rules embedded
in agents and rules embedded in the institutional environment. Both agents
and institutions are, in this sense, carriers of rules.
The complexity of an economic system C(R) is distributed across the com-
plexity of agents C(A
R
) and the complexity of the environment of institutions
C(I
R
) as such:
C(R) = C(A
R
) + C(I
R
)
Agent rules thus refers to the cognitive, behavioral, and interactive capabilities
and actions of an economic agent. Economic agents are therefore rule-using
and rule-making carriers (Dopfer, 2004; Potts, 2003). Examples of agent rules
include skills, habits, routines, capabilities, and other instances of generic
rule-behavior. These rules will have a measure of complexity.
Institutional rules, in turn, are the artifcial (or artifactual) environment of the
economic agent, and consist of the set of meso rules the agent is connected to.
Examples include contracts, capital, commodities, organization, institutions,
and so forth. These rules will also have a measure of complexity. Generic
economic rules are therefore variously carried by agent and institutional rules
in some convex combination. This maps the complexity measure we shall
call an isocomplexity function, or isocomp.
The allocation of complexity across rule-carriers assumes a 2-d rule-space
we call carrier space. Any system of rules R will be distributed between
agents (A
R
) and institutional environment (I
R
).
8
Consider Figure 2 in which
the carrier dimensions of the agent and the environment are represented along
orthogonal axes of increasing rule complexity. Each point in generic carrier-
space represents a hybrid generic rule-system across agents and institutions
at varying levels of carrier complexity.
9

The generic nature of carrier-space is conceptualized as such. Point A has
high complexity in the institutional environment and low complexity of
agents. This might be an economic system in which there is a high degree
of organization in the environment, with each agent specialized, but not
very complex otherwise. A fascist or dictatorial state would be an example,
as would a beehive. Point B has the same overall systemic complexity in
the generic rule (i.e., has the same problem-solving capabilitiesbut this is
achieved with less institutional complexity compensated for by greater com-
0 Potts, Morrson, & Clark
plexity in agents capabilities (learned rules)). This might be represented by
a well-educated population in a more rudimentary environment (e.g., J apan
after WWII, or Russia after the collapse of communism). Point C continues
along this overall complexity invariance function, and might represent highly
capable agents in a very primitive environment (as in the reality TV show
Survivor, or Western businesses expanding into emerging markets). Point D
corresponds to simple agents in a simple environment (human prehistory)
and point E corresponds to complex agents in a complex environment (e.g.,
a modern frst-world economy). Points F and G correspond to mostly unin-
habited possibilities.
A map of complexity equivalent rule substitution in carrier space is called
an isocomplexity function. For example, contour lines of isocomplexity are
traced in meso space in the C(R) level of complexity in ABC and in the C(R)
level of complexity in FEG in Figure 2. The same measure of aggregate
complexity can be achieved in various ways: either with highly complex
agents in a simple environment or simple agents in a highly complex envi-
Figure2.Thecarryingcapacityofrule-systemsbythecomplexityofagents
and environment is mapped in meso space. ABC and FEG are a convex
maps(isocompcurves)ofconstantcomplexityinmesorulesC(R)andC(R)
allocated over two classes of rule carriers C(I
R
) and C(A
R
). Isocomp curves
have similar properties to indifference curves.

High
Agent complexity C(A)
Low
High
Environment
complexity
C(I)
A
B
C
D
E
Isocomp curveC(R)
G
F
Isocomp curveC(R)
ronment; or some convex combination of the two. An isocomp represents
the distribution of carrier-equivalent computational complexity about a meso
unit that we presume to be convex in its equilibrium distribution. Convex-
ity and complexity are thus aligned as equivalent statements of the same
static/dynamic point.
A shift along an isocomp represents a change in the relative effciency of
carriers of a rule, and so leading to a rule being carried more by institu-
tions rather than agents, as in Figure 2 with a shift from C to A, or E to F.
This represents the process of institutional embedding. By substituting rule
complexity from the agent to the environment enables simpler agents to
solve the same problems as they embed rules deeper in the environment.
The entrepreneur-led processes of market-making or infrastructure building
are classic examples of this, as is the role of mass advertising or education.
A shift along an isocomp is complexity invariant at the system level, but
re-allocates complexity from one class of carrier to another. This will invari-
ably be accompanied by structural change at the agent, frm, industry, and
economy-wide level.
A shift in the isocomp, say from the locus of points though ABC to those
running through FEG in Figure 2, represents an increase in the aggregate
complexity of the rule-system. Such emergent complexity is, in fact, a neces-
sary process in generic economic evolution as mandated by the second law
of thermodynamics applied to an open system (Raine et al., 2006). Increased
energy throughput is achieved through increased complexity. But there are
many ways this can happen and without further consideration it is unclear
what the distributive effect will actually be on the generic allocation of com-
plexity between agents and institutions.
Three possibilities from point B are previously illustrated. An evolutionary
meso trajectory B to E preserves the distribution of complexity, while a trajec-
tory from B to F represents an institution-dominated growth of complexity,
and one from B to G represents an agent-dominated growth of complexity.
But what will happen is yet undetermined. All we can say is that a shift in an
isocomp, say from C(R) C(R), represents an increase in the complexity
of the aggregate rule system in consequence of the growth of opportunities
created by the growth of knowledge.
A shift along an isocomp curve is meso complexity preserving. A shift in the
isocomp curve corresponds to increased overall energy-transformation and/or
information processing or problem solving by an increase in the complexity
of the generic rules of both agents and institutions. The growth of knowledge
and the growth of economic systems are but two sides of the same problem
of the allocation of complexity.
10
The normal direction of economic evolu-
tion is to increase both the complexity of institutions and agents as an overall
increase in the energy or information throughput of the system. This will
tend to correspond to an increase in the state of knowledge of the economic
system. In general, the dynamical path of economic evolution can be traced
out in this meso space.
But, without some notion of relative costs or prices of complexity under
different allocative conditions, a generic trajectory is undetermined. The
energy-complexity framework tells us the general direction upon which we
should expect things to happen, but there is as yet no choice theory in this
argument. For that, we need to consider the implications of a budget constraint
on the allocation of complexity.

The.Relative.Price.of.Complexity
We may introduce relative prices, or relative costs,
11
into this framework
by considering the relative price of carrying a generic rule as considered
variously in terms of an agent or in terms of the agents environment. As
argued above, the existence condition of a generic rule is that it is always
able to be substituted between agent and environment, and vice versa. We
now suppose that the equilibrium outcome will be in proportion to the cost
(relative generic and operational price) of embedding complexity in the agent
or environment respectively.
The relative price of complexity between institutions and agents is given
by the budget line C(I)C(A). The slope of the budget line is given by the
relative costs of carrying complexity (P
I
/P
A
). A steep budget line means that
institutional carrier costs are relatively lower than agent carrier costs, making
it better to carry complexity of rules in institutions rather than in agents. A
fatter budget line means that agents carry the rule relatively more effciently
than institutions. A fall in the absolute cost of complexity, in both agents and
institutions, shifts the budget line outward. An increase in the absolute cost
of complexity shifts the budget line inwards. Figure 3 illustrates a shift in
the carrying costs of complexity to make agent carriers relatively cheaper
than institutions.
A pivot in the budget line as illustrated in Figure 3 might be caused by
a change in the relative effciency (or effcacy) of rule-carriers, but a shift
outward of the budget line as in Figure 4 would be caused by an absolute
change in the effciency and effcacy of rule carriers. Both sorts of change
are induced by new rules, and what specifcally happens will depend on the
particular character of the novel generic rule and its relation to other rules. For
example, the introduction of a new generic rule in the form of an improvement
in technology, or market infrastructure, or agent skills will probably change
both the relative costs of embedding and the absolute costs of embedding.
Nevertheless, our analytical framework separates the two components of
relative and absolute price changes. An absolute shift corresponds to a new
rule enabling a higher level of aggregate complexity overall, and therefore
energy throughput, but without changing the underlying carrying costs. This
defnes an outward shift of the budget constraint to a higher isocomp curve.
A suitable example would be the introduction of PCs or mobile phones, both
Figure 3. A shift in the budget line for carrying complexity. A fall in the
complexity of institutions is compensated for by a rise in the complexity of
agents
PA
C(A) '
C(I) '
B
0 A0 I
C(I)
A
Complexity of
Institutions
Complexity of Agents
0
Isocomp curve
C(A)
of which require more complex institutional support, and, in their early days
at least, increased agent complexity.
The set of different possible trajectories an evolving economic system may
trace out can therefore be analyzed in terms of different relative costs of
carrying complexity in agents and institutions. In Figure 4, we decompose
the comparative statics of this process into the conventional microeconomic
notions of income and substitution effects. Beginning at point E, a fall in the
carrying cost of agent complexity induces a shift of the equilibrium to point
F. The substitution effect is represented by ED and the income effect from
DF. The decomposition technique of income and substitution analysis may
provide analytical insight in a number of ways. For example, by indicating
the full magnitude of the uncompensated substitution (ED) as distinguished
from the relative shift in overall complexity, we might better understand
how to structure industrial transitions and the sort of agent and institutional
support required.
This same sort of analytical reasoning may also be applied to the manage-
ment of change in businesses, as about the balance between centralizing
Figure4.Incomeandsubstitutioneffectstracedoutbydifferentrelativecosts
of carrying complexity. A reduction in the relative carrying cost of complexity
in agents induces a shift in overall complexity. This can be decomposed into
an income effect and a substitution effect.
PA
Income
effect
Substitution
effect
C(A) '
C(I) '
F
E
C(I)
D
Complexity
of Institutions
0
C(A)
and decentralizing such a process of rule substitution, as say the process
of adaptation to a new market or technology. Substitution effects therefore
measure the cost of change, and income (or wealth) effects measure the gain
from change. Of course, without formal elucidation of the analytical mean-
ing of a complexity measure, these concepts are still somewhat conjectural.
But they do hint at the possibility of a complexity-based welfare analysis of
evolutionary economic processes.
There are other aspects of this framework we might explore. For example,
Figure 5 illustrates how we may represent the dynamical path of a complex
evolving system by constructing this as a meso trajectory of relatively con-
stant complexity embedding. This forms, as mapped through 0BD, an evo-
lutionary contract curve of system coordination.
12
The connection of this
to the Schumpeterian notion of a trajectory should be obvious to scholars of
evolutionary economics. Allocation of complexity theory therefore offers a
method for analyzing evolutionary economic processes at the level of rule
substitution between classes of carriers as a function of the relative price of
doing so. In this way, rule complexity is allocated between carrier dimensions
Figure 5. A shift out in the budget line for carrying complexity along a meso
trajectory. The introduction of a new meso rule that is agent-institution
neutral corresponds to a shift out in the budget line and therefore to a new
tangency condition on a higher isocomp (from B to D).
PA
C(I)
D
C'
B
Complexity of
Institutions
0
Isocomp curve
Meso complexity
trajectory
C
C(A)
as a function of the relative price of embedding said rule complexity into each
carrier dimension. Complexity goes where incentives tell it to go.
Conclusion
This chapter has sought to advance the theory of complex adaptive systems
through consideration of a prime open-system economic dimension: namely,
the allocation of generic rule complexity over different classes of carriers.
What we argued was essentially that a standard neoclassical economic analy-
sis of relative prices in the context of continuous substitution can provide
a useful method for the analysis of an important economic dimension the
behavior of an intelligent complex adaptive system, namely the allocation of
generic complexity over carrier-space. We have proposed a comparative static
methodology for analysis of economic evolution that, following Dopfer et al.
(2004) and Dopfer et al. (2007), is conceptualized as a generic rule process.
The basic ontological building block of this view is a meso unit: a generic
rule and its population of carriers. Economic evolution is defned within this
framework as an increase in the rule complexity of an economic system. But,
in our view, the essential economic question is: complexity where? When
knowledge grows, where does it accumulate?
Allocation of complexity theory is a way of analyzing the allocation problem
inherent in this set-up, namely between changes in complexity in agents and
changes in complexity in the economic environment in terms of continuous
substitution along an isocomplexity function in terms of the relative price of
adoption and retention. According to our theory, then, when the relative price
of adoption and retention changes, so the equilibrium allocation of complex-
ity will shift. And while we emphasize the foundational importance of this
analytic concept to the study of economic evolution, we also wish to draw
attention to the signifcance of this generic allocation concept in the evolution
of all complex adaptive systems, intelligent ones especially: for what is intel-
ligence, but the proper balance of dynamic engagement? J ust as evolutionary
biology foists a genotype/phenotype distinction on evolutionary social science,
evolutionary economics has a generic distinction between classes of carriers
of generic rules. Variously, each new idea or technology is best carried in
the human mind, or as a pure technology, or some convex combination. The
analytic difference, we suggest, is the allocation of complexity.
References
Allen, P. (1998). Evolving complexity in social science. In G. Altman, & W.
Koch (Eds.), Systems: New perspectives for the human sciences. Berlin:
Waltyer de Gruyter.
Dopfer, K. (2004). The economic agent as rule maker and rule user: Homo
Sapiens Oeconomicus. Journal of Evolutionary Economics, 14, 177-
195.
Dopfer, K., & Potts, J . (2007). The general theory of economic evolution.
London: Routledge.
Dopfer, K., & Potts, J . (2004b). Evolutionary realism: A new ontology for
economics. Journal of Economic Methodology, 11, 195-212.
Dopfer, K., Foster, J ., & Potts, J . (2004a). Micro meso macro. Journal of
Evolutionary Economics, 14, 263-79.
Earl, P. E, & Potts, J . (2004). Market for preferences. Cambridge Journal of
Economics, 28, 619-33.
Earl, P. E., & Potts, J . (2001). Latent demand and the browsing shopper.
Managerial and Decision Economics, 21, 111-22.
Foster, J . (2006). From simplistic to complex systems in economics. Cam-
bridge Journal of Economics, 29(4), 873-92.
Foster, J ., Ramlogan, R., & Metcalfe, J . S. (2006). Adaptive economic growth.
Cambridge Journal of Economics, 30(1), 7-32.
Hayek, F. (1960). The constitution of liberty. Chicago: University of Chicago
Press.
Loasby, B. (1999). Knowledge, evolution, and institutions in economics.
Routledge: London.
Mirowski, P. (2007). Markets come to bits: Evolution, computation and
markomata in economic science. Journal of Economic Behavior and
Organization, 63(2), 209-42.
Ormerod, P. (2006). Why most things fail: Evolution, extinction, and econom-
ics. London: Faber and Faber.
Potts, J . (2003). Toward an evolutionary theory of homo economicus: The
concept of universal nomadism. In J . Laurent (Ed.), Evolutionary eco-
nomics and human nature. Edward Elgar: Cheltenham (ch 8).
Potts, J . (2001). Knowledge and markets. Journal of Evolutionary Econom-
ics, 11, 413-31.
Potts, J . (2000). The new evolutionary microeconomics: Complexity, compe-
tence, and adaptive behaviour. Edward Elgar: Cheltenham
Potts, J ., & Morrison, K. (2007). Meso comes to markets. Journal of Eco-
nomic Behavior and Organization, 63(2), 307-12.
Prigogine, I. (2005). The rediscovery of value and the opening of economics.
In K. Dopfer (Ed.), Evolutionary foundations of economics. Cambridge:
Raine, A., Foster, J ., & Potts, J . (2006). The new entropy law and the eco-
nomic process. Ecological Complexity, 3(4), 354-60.
Schneider, E., & Kay, J . (1994). Life as a manifestation of the second law of
thermodynamics. Mathematical Computational Modeling, 19, 25-48.
Shackle, G. L. S. (1972). Epistemics and economics: A critique of economic
doctrines. Cambridge: Cambridge University Press.
Simon, H. (1969). The architecture of complexity. Proceedings of the Ameri-
can Philosophical Society (Vol. 106, pp. 476-82).
Endnotes
1
Special thanks to participants at Organizations, innovation and com-
plexity New perspectives on the knowledge economy, NEXUSCRIC
Workshop, University of Manchester, 910 Sept 2004, and especially
B Rosewell and P Ormerod, as well as those at Shifting boundaries:
Governance, competence and economic organization in the knowledge
economy, Bristol Business School, 23 Sept 2004, especially N Foss, N
Kay and B Nooteboom. Also, special thanks to K Bruce, R Ramlogan,
J S Metcalfe, K Dopfer, and J Foster.
2
The views expressed in this paper are his own and do not necessarily
refect the opinions of Suncorp.
3
Earl and Potts (2004) have recently illustrated how choice itself can
be considered in this way when decision complexity leads boundedly
rational consumers to outsource decision rules to market environments.
Their market for preferences concept can be viewed as an allocation
of rule complexity problem, when the rules are inputs into a decision
heuristic conceived along an isocomp. See also Earl and Potts (2001)
on the allocation of complexity with respect to the design of shopping
environments.
4
One might reasonably suppose that the answer to this question is trans-
actions costs, in the Ronald Coase or Oliver Williamson. Isocomplexity
theory is more aligned with Coases original version about alternative
coordinating institutions rather than Williamsons development of the
idea into statements about bargaining games, etc. We suggest that iso-
complexity theory provides a more general theory of this process, by
taking into account the costs of embedding rules in a system, as well
as the ongoing costs of using them. Isocomplexity theory deals with
systems costs generally, an of which transaction costs are a subset.
5
This is also called exosomatic evolution. It is a broader concept than
what Richard Dawkins calls memes. See Dopfer and Potts (2007) for
general discussion.
6
A further way of treating complexity, recently proposed by Foster (2006),
is with the notion of an economic system as being 4
th
order complex
(dissipative physiochemical systems are 1
st
order complex, biological
systems are 2
nd
order complex, social systems are 3
rd
order complex).
The emergence of higher order complexity turns on the emergence of
new ways of acquiring and using energy through ever more complex
systems of knowledge and interaction. An economic system is 4
th
order
complex in that agents must not only form mental models of the environ-
ment (3
rd
order) but also of each other mental models. This is apparent
in such things as contracts and other market institutions.
7
See Ormerod (2006).
8
The algebraic properties of meso space derive from the set-theoretic
conception of the carrier dimensions. It is unclear to us whether it is
reasonable to assume these to be real valued. It would certainly be
convenient. At the risk of unhinging the analysis before it starts, let us
make that a point for mooting.
9
We immediately rule out corner solutions by assuming that there is no
such thing as an economic system purely made of agents, and similarly,
no such system purely made of environment. An economic system is
a system of economic systems, and each economic system consists of
a non-zero combination of both agents and institutions. All economic
0 Potts, Morrson, & Clark
rules are embedded in both agents and institutions and are fungible
between.
10
See Hayek (1960), Shackle (1972), Loasby (1999).
11
Because each point on an isocomp corresponds to the steady state of a
general equilibrium under competitive conditions (and so extensions to
this framework arrive under conditions of imperfect competition), then
each set of relative prices, under duality conditions, also corresponds
to the set of relative costs.
12
This raises a number of analytic questions. What, for example, is the
nature of the factors that determine relative cost of embedding, or the
manner in which complexity is structured in an aggregate system? Does
the network classifcation of the system matter (i.e., does the shape of
the isocomp depend upon whether we are dealing with a small world or
scale-free network?) Does the distribution of complexity over a popula-
tion of agents matter?
About the Contrbutors
About t he Cont r i but or s
Ang.Yang joined the Division of Land and Water at the Commonwealth
Scientifc and Industrial Research Organisation (CSIRO) in 2007. Yang
holds a PhD in computer science (UNSW, Australia), a MInfoSc in informa-
tion systems (Massey University, New Zealand), an MSc in environmental
geography (Nanjing University, China), and a BSc in ecology & environ-
mental biology (Ocean University of China, China). His current research
interests include complex adaptive systems, multi-agent systems, modelling
and simulation, evolutionary computation, network theory, and Web-based
intelligent systems.
Yin.Shan is with Medicare Australia as a senior review assessment offcer
working on data mining and machine learning applications on medical data.
He previously worked as a scientifc programmer in the Australian National
University and a postdoctoral research fellow in the University of New
South Wales, Australia. He received his PhD in computer science from the
University of New South Wales (2005) and his MSc and BSc in computer
science from Wuhan University, China (1999 and 1996, respectively). His
main interests are evolutionary computation, in particular genetic program-
ming and its applications.
.......................................................*..*..*
Hayward.R..Alker graduated in 1959 from M.I.T. in mathematics and re-
ceived his PhD in political science from Yale University (1963). After serving
on the Yale and M.I.T. faculties in political science, he is now a professor of
international relations at the University of Southern California and an adjunct
professor at the Watson Institute of International Studies, Brown University.
He has served on the executive committee of the International Social Science
Council as chair of the Mathematical Social Sciences Board and president
of the International Studies Association.
Gordon.L..Bonser obtained a bachelor of science degree in natural resource
planning from Humboldt State University, Arcata, California (1974). Gordon
taught geographic information systems (GIS) at Humboldt State University,
Department of Natural Resources and Sciences (http://www.humboldt.
edu/~cnrs/departments/enrs.html) His research specialized in the feld of
multi-spectral analysis of images acquired using space-based platforms such
as LANDSAT for the purpose of ecosystem-level vegetation mapping and
modeling. Projects included analysis of human impacts on salmon habitat in
a large watershed in Northern California (1995) and an analysis of changes
in fre frequency and intensity on the Snake Rive Plains of Idaho as a result
of rainfall, introduced Asian plants, and land use (1996). Presently he is
semi-retired.
Joseph.Clark is a PhD candidate in the School of Economics at the Univer-
sity of Queensland. He is also a research economist at Suncorp Pty. Ltd. The
views expressed in this chapter are his own and do not necessarily refect
the opinions of Suncorp.
David.Cornforth is senior lecturer at the University of New South Wales.
He has been an educator for twenty years, including teaching a variety of
discipline areas at four different universities. He has worked as a program-
mer and data analyst in the automotive, tourist, fnancial, health, water sup-
ply, and electricity supply industries. He has been involved in research in
complex systems for the last seven years, and has published a large number
of research articles on this and related topics.
Jean-Louis.Dessalles is associate professor at ParisTech, cole Nationale
Suprieure des Tlcommunications. His domain of research is about the
cognitive modelling of language and about the emergence of stable com-
munication in a population of selfsh agents. He authored a book on the
emergence of language, now translated in English (Why We Yalk, Oxford
University Press, 2007).
Guszti.Eiben is professor and head of the computational intelligence group
in the department of computer science at the Vrije Universiteit in Amsterdam,
the Netherlands. His main research interest is evolutionary computation.
Evolutionary computation is the binding theme in much of Gusztis academic
research that can be catalogued under natural computing, or soft computing,
with applications in optimization, data mining, artifcial life, and artifcial
societies. Specifc research subjects are multi-parent recombination, constraint
handling and self-calibrating algorithms, the effect of communication in
Artifcial Life, and modeling evolving societies.
Jacques.Ferber is professor of computer sciences at Montpellier. His domain
of research is about multiagent systems (MAS), distributed systems, multia-
gent simulation, component-based and agent-based software engineering. He
started to work in MAS in 1987 and is considered as a pioneer of this feld in
France. His introductory book on MAS and distributed artifcial intelligence
has been published in French, English, and German. Member of several edi-
torial boards and program committee (conferences and workshops), Ferber
is the co-author with Olivier Gutknecht and Fabien Michel of the MadKit
multiagent platform.
David.G..Green is professor of information technology at Monash University.
In the course of thirty-fve years of research on complexity he has investi-
gated problems as diverse as forest ecology, proteins, and social networks.
He proved the universality of networks in complex systems and is author of
9 books and more than 160 research articles on complexity theory, evolution-
ary computing, and multi-agent systems. An early pioneer of the World Wide
Web as a distributed information resource, he also played a leading role in
national and international efforts to create a coordinated Web of information
sources about the worlds biodiversity and environments.
Thomy.Nilsson, a BSc 63 from Rensselaer, provided a physics background
and, under Kandel, experimental psychology. At Columbia, Kietzman intro-
duced the latest research technologies. For MSc 68 & PhD 72 with Thomas
Nelson at Alberta, Thomy Nilsson developed vision laboratories and later as
research associate expanded into ergonomics; Wilson introduced artifcial
intelligence. Bourassa at Portlands Neurophysiology Institute and Donchin
at Illinois broadened Thomy Nilssons technical and physiological skills. A
professorship at UPEI added research on aging. Teaching inspired statistics
and research methods books. Thomy Nilssons vision research focuses on
color, temporal factors, legibility, and memory; discovered visual temporal
tuning, cool effect, memory retrieval characteristics, pinhead mirror, and
chromatic legibility. Thomy Nilsson is now an emeritus professor.
Geoff. Nitschke is a research assistant working within the computational
intelligence group in the department of computer science at the Vrije Univer-
siteit in Amsterdam, the Netherlands. Nitschkes research interests include
evolutionary computation, neural networks, artifcial life, and collective
behavior systems. The research issue being addressed is how concepts such
as self-organization, evolution, and learning can be used in formulating bio-
logically inspired design principles for collective behavior systems. The goal
is to apply such biologically inspired design principles as problem solving
methods in collective behavior systems. Examples of such collective behav-
ior systems include collective gathering, reconnaissance, and construction
in multi-robot systems.
Kate. Morrison is director of Volterra Pacifc Pty. Ltd., a Brisbane-based
economics consultancy that specializes in complexity based analysis.
Pascal.Perez is currently seconded by his French research agency CIRAD
to the Australian National University (ANU) in Canberra (Australia). He is
the convenor of the Human Ecosystems Modelling with Agents (HEMA)
international network. Agronomist by training, his most recent work focuses
on human ecosystems modelling. He has developed projects in northern
Thailand, Indonesia, and Micronesia. He is currently teaching agent based
modelling at the ANU.
Denis. Phan is post and telecommunications administrator, doctor in eco-
nomics, and senior researcher in economics and social sciences at the CNRS.
His past research interests are in the economics of telecommunications, the
software industry, and regulation. He has published papers and various books
in this feld. Since 2000, his main research interests have been cognitive and
agent-based computational economics. He is the co-author with Antoine
Beugnard of the Moduleco simulation Framework, now integrated into the
Madkit Platform as a plug-in. He has organized several researcher schools
on complex systems and MAS and published with Frederic Amblard an
introductory book on MAS modeling and simulation for Human and Social
Sciences in French, to be traduced in English for The Bardwell Press.
Jason.Potts is senior lecturer at the School of Economics & ARC Centre for
Complex Systems, University of Queensland, Australia. He is currently on
secondment as principal research fellow at the ARC Centre of Excellence for
Creative Industries and Innovation, Queensland University of Technology,
Australia.
Grigorii.S..Pushnoi has a masters degree in physics from the St. Peters-
burg State University, Russia. He is an investigator of Complex Evolving
Systems. Working as a manager of business organizations in St. Petersburg,
he observed the dynamics of those businesses development: why some
organizations are successful, while others fail or disintegrate. This valuable
experience helped him to fnd a new vision of the evolutionary processes. He
developed a unique top-bottom technique of the complex adaptive systems
modeling: the Method of Systems Potential: (MSP). His report for the 21st
International Conference of the System Dynamics Society is included in the
System Dynamics online course (Politechnic University of Catalonia; Terrassa,
Spain). His papers dedicated to applications of MSP in Econophysics and
Business Cycle Theory are published in the Proceedings of the International
A.Bogdanov Institute in Russia. He moderates the interdisciplinary scientifc
forum: Socintegrum, which is dedicated to the problems of civilizations
development and their solutions.
Kurt.A.. Richardson is the associate director for the ISCE Group and is
director of ISCE Publishing, a recently founded publishing house that spe-
cializes in complexity-related publications. He has a BSc(hons) in physics
(1992), MSc in astronautics and space engineering (1993), and a PhD in
applied physics (1996). Kurts current research interests include the philo-
sophical implications of assuming that everything we observe is the result
of underlying complex processes, the relationship between structure and
function, analytical frameworks for intervention design, and robust meth-
ods of reducing complexity, which have resulted in the publication of 25+
journal papers and book chapters. He is the managing/production editor for
the International Journal Emergence: Complexity & Organization and is on
the review board for the journals Systemic Practice and Action Research,
and Systems Research and Behavioral Science, and has performed adhoc
reviewing for Journal of Artifcial Societies and Social Simulation and
Nonlinear Dynamics, Psychology, and the Life Sciences. Kurt is the editor
of the recently published Managing Organizational Complexity: Philosophy,
Theory, Practice (Information Age Publishing, 2005) and is co-editor of the
forthcoming books Complexity and Policy Analysis: Decision Making in an
Interconnected World (due August 2007) and Complexity and Knowledge
Management: Understanding the Role of Knowledge in the Management of
Social Networks (due December 2007).
Martijn.Schut is an assistant professor working within the computational
intelligence group in the department of computer science at the Vrije Uni-
versiteit in Amsterdam, the Netherlands. Martijns research concerns the
understanding of the behavior and self-organization of complex systems. That
is, systems in which the interaction of the components is not simply reduc-
ible to the properties of the components. Martijns research addresses how
systems of very many independent computational agents should cooperate in
order to process information and achieve their goals, in a way that is effcient,
self-optimizing, adaptive, and robust in the face of damage or attack.
Russell.Standish is a computational scientist and principal of High Perfor-
mance Coders, a computational science consulting practice based in Sydney,
Australia. He has extensive contacts with academia, with a visiting associ-
ate professorship at the University of New South Wales, and also an active
research program in Complex Systems, focussing on evolutionary systems
using artifcial life techniques. He has recently published a book Theory of
Nothing that applies the notions of complexity and information to tackle
age-old ontological questions.
Hrafn.Thorri.Thrisson has been studying natural and artifcial creative
systems for several years. His research has earned signifcant recognition,
including Icelands National Young Scientists Award and a nomination for
the Icelandic Presidential Innovation Award. He studies computer science at
Reykjavik University and is a member of the Center for Analysis and Design
of Intelligent Agents. Thrisson is the founder of Icelands frst and only A.I.
association, The Icelandic Society for Intelligence Research, and currently
resides as its president.
Steven E. Wallis received his PhD from Fielding Graduate University in 2006.
His academic work focuses on -Ytheory of theory where he is developing
insights and tools to support scholars as they create increasingly effective
theory. Walliss interdisciplinary approach spans the social sciences. In addi-
tion to complexity theory, a few of his interests include collaborative human
systems, knowledge management, and organizational theory. Dr. Wallis has
ten years of experience as an independent consultant in Northern California.
There, in a variety of industries, he supports consultants, trainers, and lead-
ers on issues related to collaboration, communication, succession planning,
creativity, and knowledge management.
Index
I ndex
Symbols
2-dimensional arrays 172
4-dimensional tensor calculus 116
4-quadrant approach 280
4-quadrant point of view emergence 285
A
agent-based modelling (ABM) 26
ABM bottom-up modeling 26
actual simulation 235
adaptation 12
adaptive activity 30
adaptive activity, frst postulate of 44
adaptive activity, second postulate of 44
adaptive multi-agent modeling 313
adaptive systems, central processing 159
adjacency matrix 109
agent-based computational economics
(ACE) 256
agent-based computational social science
255
agents 13
aggregation 99
algorithmic complexity 114, 276
algorithmic complexity, defnition 123
algorithmic information complexity 117
algorithmic information theory 114
allocation of complexity 332
Anthropic Principle 117
Anthropic Principle, defnition 123
arcs 107
artifact 129
artifcial chemistry 98
artifcial cognitive system 136
artifcial complex adaptive systems
98, 216
artifcial intelligence (AI) 127
artifcial life (ALife) 135
artifcial life systems 216
artifcial species 136
artifcial systems 75, 91
atomistic concepts 16
autopoiesis 131
B
behavioral specialization 218, 220, 221
bifurcation 192
biological collective behavior systems 216
biologically inspired design principles 216
biological taxonomy 87
Index
biological taxonomy, hierarchy of 87
Boolean networks 187
Boolean networks, dynamics 189
Boolean networks, information barriers
187
Boolean networks, modularization 202
Boolean networks, structure 189
bottleneck problem, solution 166
building blocks 84
C
carrier space 339
CAS, adaptive potential of 29
CAS, conditions of realisation 29
CAS-modelling 39
CAS-modelling, MSP-platform 42
CAS agents 28
CAS modelling 35
CAS theory 1, 2
CAS theory, concepts of 9
CAS theory, conceptual concepts 2
CAS theory, reductive study of 4
CAS theory, studies of 1
causal concepts, relationship between 14
causal propositions 13
causal relationship 13
cellular automata 133, 135
cellular automaton 119, 275
cellular automaton, defnition 123
church-turing thesis, defnition 123
classifer system, defnition 123
cliocide 300, 304
co-causal propositions 1
co-selection 101
cognition 131
collective behavior methods for specializa-
tion 221
collective behavior systems 215, 217
collective behavior systems, biologically
inspired principles 215
collective behavior tasks and specializa-
tion 226
collective construction 228
collective gathering 226
collective resource allocation 230
collective resource distribution 230
complex adaptive system, construction of
37
complex adaptive systems (CAS)
1, 2, 26, 27, 35, 76, 77, 215
complex adaptive systems, modularity and
75
complex adaptive systems modelling 26
complex adaptive systems models 302
complexity 75, 187
complexity (quality), defnition 123
complexity (quantity), defnition 123
complexity, allocation of 331, 338
complexity, concept of 105
complexity, context dependence 107
complexity, defnition of 105
complexity, graph theoretic measures of
107
complexity, information 110
complexity, relative price of 342
complexity as a quality 106, 117
complexity as a quantity 106
complexity economics 331
complexity in economic systems 331
complexity theory 2
complex life forms 159
complex networks, dynamic robustness
204
complex rules 331
complex systems 117, 334
complex systems, adaptive nature 77
complex systems, robustness 187
component concepts 2
computational complexity 115
computational complexity, defnition 123
concept of refexive 9
connected graph 80
connectivity 305
connectivity of a graph 108
constant change 12
contextual archetypes 194
control network 140
convergence 165, 170
convergent-divergent branching networks
166
convergent-divergent comparison networks
180
convergent-divergent decoding 164
0 Index
convergent-divergent encoding 164
convergent-divergent networks 159, 163
convergent-divergent neural network 172
convergent-divergent systems 165, 179
cooperative co-evolutionary genetic algo-
rithms 222
cooperative co-evolution methods 222
cooperative transportation tasks 238
cortical neurons 165
cost of change 345
creative artifacts 129
creativity 150
creativity, emergence of 126
creativity, simulation approach 132
Crude Look at the Whole (CLAW) Work-
shop 313
cycle 80
cyclical dynamics 27
cyclomatic number of a graph 109
D
Darwinism 42
data classifcation 85
decision-making processes 187
decision cortex 138
decision tree 86
density 118
destination map 192
DEWEY catalogue system 78
diagrammatic 4-quadrant approach 256
diameter of a graph 80
digital literacy 1, 26, 105, 126, 159, 18
7, 215, 255, 300
dimensional analysis 16
directed graph 80
distance-based methods 86
divergence 165
division of labor methods 222
dominant zone similarity 139
downward/backward causation 255
downward causation 258, 265
dynamical robustness 194
dynamic core, emergence 198
dynamic equilibrium 12
dynamic robustness of complex networks
204
dynamics 27
dynamics cores 195
E
Earth-Moon-Sun system 115
economic systems 331
edge of chaos 69
edge of chaos (EOC) 13
edges 107
effective complexity 114
Einsteins General Theory of Relativity
116
embryogenesis 89
emergence 117, 256
emergence, conceptual issues of 257
emergence, defnition 123
emergence, defnitions 271
emergence, varieties of 261
emergence as a complexity drop 275
emergence of a dynamic core 198
emergence of creativity 126
emergences in complex systems 272
emergent behavioral specialization 217
emergent behavior design methodologies
217
emergent dynamic structure 198
emergentism 262
emergent properties 27
emergent properties, defnition 123
emergent specialization 215, 220
emergent specialization, heterogeneous
approaches 224
emergent specialization, homogeneous ap-
proaches 224
emergent specialization design methodolo-
gies 242
energy payoff 139
enforced sub-populations (ESP) 223
entail, defnition 123
entropy 118
entropy, defnition 123
episodic memory 136, 139
equilibrium-based analysis 332
equivalence class, defnition 124
equivalence relation 84
event-data coding practices 304
Index
event data 302, 305
event data research 310
evolution 217
evolutionary algorithms 222
evolutionary economics 333
external observer 272
eyeless gene 87
F
face validity 303
fault tolerance 95
frst-order strategy 210
frst law of thermodynamics 118
frst order functionality 207
fxed evolutionary parameters 53
fruit fies 87
fully connected graph 80
functional space 194
G
Game of Life, defnition 124
game theory methods 222
genetic algorithm (GA) 275
genetic algorithms 98
genetic modules 87
genotype 194
global dynamics 211
Gdel Incompleteness Theorem, defnition
124
graph 78, 107
graph theoretic measures 108
graph theory 79
graph theory, defnition 124
grounded theory 16
growth hormone 89
H
halting problem 113
hard-core assumptions 2
hierarchical structure 78
hierarchies 75
hierarchy 82
hierarchy, defnition 78
Hiroshima ontologies 300
historicity 305
historicity, modeling 315
horizontal emergence 202
hormone 89
human civilization 128
human creative capabilities 128
human creativity 129
humanistic history 305
I
ideal gas 117
ideal gas law 118
image matching 139
implicit goal 151
information 187
information, defnition 124
information, reduction of 187
information age 187
information barriers 187, 202
information bottlenecks 159
information conserving loops 199
information fows 187
information hiding 92
information loops 195
information processing 181
information theoretic graph complexity
114
inheritance 93
input gate boxes 140
input location 165
input strength 166
insanity 150
integrated development environment (IDE)
91
intelligence 126
intelligent complex adaptive systems
(ICAS) theory 1, 2
inter-molecular interaction 35
interactions 13
intrinsic emergence 275
K
Kolmogorov complexity 114
L
labor methods, division of 222
Lamarcks First Law 43
Lamarcks laws 36
Index
Lamarcks Second Law 43
Lamarckism 42
language hierarchy 91
lateral pressure theory 313
learning 217
leaves 82
LEGO robot 136
length of a path 80
Leyton machine 276
Linear Array Model 167
link degree, defnition 124
links 78
logical couplings 210
logical depth 115
logical depth, defnition 124
logicality 150
loop types 197
M
M-Strong emergence 274
macro-indices 27
macroscopic mechanisms of adaptation 29
MAM-MSP interrelation 35
MAM-platform 36
method of systems potential (MSP)
26, 29, 42
micro code 91
modular graph 79
modularity 75
modularity, adaptive processes 99
modularity, defnition 76
modularity and CAS 75
modularity and fault-fnding 98
modularity and fault tolerance 95
modularity and the Internet 95
modularity in artifcial complex adaptive
systems 98
modularity in artifcial systems 91
modularity in electronic appliances 97
modularity in embryogenesis 89
modularity in language 89
modularity in manufacturing 96
modularity in mathematics 84
modularity in natural systems 86
modularity in plants 88
modularity in social structure 90
modularity in software engineering 91
modularity in the genome 87
modularization 195
modularization in Boolean networks 202
modularization in complex networks 196
modules 75, 84
morphologically specialized robots 218
morphological specialization 218, 220
motifs 87
MSP-system, adaptive activity 44
MSP-system, density of conditions 46
MSP-system, discontinuous evolutional
cycle of 56
MSP-system, discontinuous evolutionary
cycle of 55
MSP-system, effciency 46
MSP-system, exploited potential of 44
MSP-system, second type of 67
MSP-system, temporary equilibrium states
47
MSP-systems 42
MSP-systems, frst type of 55
multi-agent computer games 233
multi-agent ecological models 312
multi-agent modelling approach (MAM)
27
multi-agent systems 280
multi-agent systems (MAS) 256
multi-robot swarm intelligence 216
multiple realizability 267
N
NAND boxes 140
natural creativity 126
network 78
network modules 83
networks, dynamical robustness 194
networks and graphs 78
networks and modularity 78
network topologies 188
neural branching 163, 179
Newtons laws of motion 118
Newtons laws of motion, defnition 124
nodes 78, 107, 197
nominal emergence 261
non-conserving (structural) information
loops, role of 201
Index
non-conserving information loops 195
non-emergent specialization 220
non-linear dynamics 2
non reductive individualism (NRI) 267
non reductive materialism 264
noosphere 280
NOR boxes 140
O
object-oriented languages 92
object-oriented paradigm 92
object-oriented programming, implications
of 94
object oriented paradigm 91
Occams Razor 116
Occams Razor, defnition 124
offdiagonal complexity 109, 110
order parameter 272
organizational theory 2
P
p-loop 194
peace and war, human history 302
peace and war, ontological refections 300
perception cortex 136, 137
percolation threshold 108
phase space portrait 192
phenotype 194
plan composer 142
positive reinforcement 101
preference-values, dynamics of 41
pressure 118
process modularisation 97
product modularisation 96
propositions 2
protective belt 2
punctuated anytime learning 236
punctuated equilibrium 27, 66
punctuated equilibrium, dynamics of 32
pursuit-evasion 236
Q
quadrants 9, 280
qualitative chaos 208
quasi-chaotic 194
quasi-chaotic attractors 194
R
random graph 80, 108
random graphs and networks 80
reduction 119, 264
reduction of information 187
refexive dimensional analysis (RDA) 16
reinforcement learning (RL) methods 229
reinforcement learning methods 223
relatedness 9
relational propositions 11
relative algorithmic complexity (RAC)
276
relaxation time 192
resource modularisation 97
response magnitude 170
RoboCup Soccer 235
robust core of theory 1
robustness 9, 187, 207
root node 82
root of the tree 82
rule-based methods 86
rule-making carriers 339
S
scale-free graphs 109
scale free distribution, defnition 124
scale free network 81
schemas 13
science of networks 188
second law of thermodynamics 119
defnition 124
segregation 99
self-organization 2, 12, 217
self-organizing components 2
semantic content 112
semantic layer 119
semantic level (or space), defnition 124
sensory information bottleneck 159
separatrices 192
set theory 85
simple graph 79
simple systems 117
simplicity, theory of 116
simulated multi-robot systems 227
simulation framework 134
six degrees of separation 109
Index
small-world network 82
small world model 108
small world networks 109
SOC-phenomenon 27, 35, 69
social intelligence 269
sociosphere 280
Solomonoff-Levin distribution 113
spanning trees 109
sparse graph 80
spatial acuity 181
spatial multiplexing 180, 181
spatial tradeoffs 181
specialization 216
specialization, suppositions of 217
specialization, types of 219
stabilizing feedback 31
state space compression 207
static boxes 140
static structure of the network 198
stick pulling 227
stochastic discontinuous cycles 33
straw man argument 120
structural attractor 198
structural complexity 276
structured programming 92
structure of theory 7
subtree 82
synchrony 100
syntactic layer 119
syntactic level (or space), defnition 124
system dynamics 215
systemic interrelation 69
system robustness 209
T
top-bottom approach 26
tatami 77
temperature 118
temporal multiplexing 181
temporal tradeoffs 181
tension index 33
theory of natural creativity 126
thermodynamic description of the gas 118
thermodynamics, frst law of 118
thermodynamics, second law of 119
tolerance 170
transients 190
tree 82
triangulation 15
Turing complete syntactic language 114
turing machine, defnition 124
U
uniform measure 113
universal macroscopic adaptive mecha-
nisms (UMAM) 29, 42
universal prior distribution 113
universal turing machine (UTM) 113
universal turing machine, defnition 124
V
vertex, degree of 80
vertices 78, 107
W
walls of constancy 195
war and peace 300
weak/strong emergence 255
weak emergence 261
weak emergent phenomenon 274
whole-part hierarchy 92
wild disjunction 267
Winner-Take-All Economics 169
Wolfram one-dimensional network of
automata 262

(Ang Yang, Yin Shan) Intelligent Complex Adaptive

Uploaded by

Copyright:

Available Formats

You might also like

(Ang Yang, Yin Shan) Intelligent Complex Adaptive

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(Ang Yang, Yin Shan) Intelligent Complex Adaptive

Uploaded by

Copyright:

Available Formats

IGIP

Acquisition Editor: Kristin Klinger

Increment in potential on account of activity.

Increment in conditions on account of activity.

Stock-fows balance equations

Stock-fows balance equations

Lamarcks frst law. (7)

are fulflled at this point. It is evident that

is not controlled by the system. The value

, thus shifting its state upwards or downwards along

gives us the following result:

< < . (27)

structure of the network of automata..

You might also like